Tasks and Methods in Applied Artificial Intelligence

Some fifty years back, a synergetic process took place which was to produce ever greater emphasis on the multi-disciplinary approach to science, relating to communication, coding, control and computation in human beings and machines. We attempt here to give a brief overview of the past evolution and future trends for some of the original concepts, mostly those pertaining to sensory coding, logical nets and neural computation to varying degrees of complexity. It is an attempt to make us all relive the excitement and the extraordinary force of the subjects together with the realms of inspiration still to be discovered in the work of the pioneers. It is also a modest contribution towards the commemoration of the Centenary of McCulloch's birth and the fiftieth anniversary of Wiener's book Cybernetics (1948)

The role of perception and action in current AI systems is analyzed and some points concerning present AI methodologies are discussed. It is argued that if we accept as a central goal for AI to build artificial systems that behave in an intelligent way in the real world, then additional evaluation criteria for intelligent systems should be used. Finally, some of these criteria are proposed.

This paper presents a new variant of the A* class algorithms for solving the known ‘one-dimensional cutting stock problem’ in a cardboard factory where the objective is to minimise the useless remaining of a continuous cardboard surface. The algorithm limits the number of nodes in the tree search by using a bound criterion based on the problem restrictions. The process of computing the solution is achieved at several stages, obtaining a complete non-optimal solution at each stage and improving the response as long as new stages are executed. The final stage returns the optimal solution. The proposed approach allows for a solution at anytime in the process resolution and also the refinement of the solution as more time is given to the algorithm. In this way, if a non-optimal solution is satisfactory enough for the factory, the process can be interrupted at that time. The computational performance of the algorithm indicates the effectiveness of the algorithm for solving practical one-dimensional cutting stock problems. Additionally, the experimental results show the important money savings achieved for the factory.

This paper is mainly dedicated to analyse the problem of discovering frequent temporal patterns in event sequences extracted from a large repository of newspapers. The proposed formalism and algorithms rely on Toodor, which is a document retrieval system that allows users to specify conditions over the structure, contents and temporal features of the stored documents. We develop in this work several algorithms for recognising frequent temporal patterns in terms of arc-consistency, which consist of discarding temporal occurrences that do not satisfy a temporal structure.

From an Al perspective, the job-shop is a constraint satisfaction problem (CSP), and many specific techniques have been developed to solve it efficiently. In this context, one may believe that generic search and CSP methods are not appropriated for this problem. In this paper, we contradict this belief. We show that generic search and CSP algorithms and heuristics can be successfully applied to job-shop problem instances that have been considered challenging by the job-shop community. In particular, we use forward checking with support-based heuristics, a combination of a generic CSP algorithm with generic heuristics. We improve this combination replacing the depth-first search strategy of forward checking by a discrepancy-based schema, a generic search strategy recently developed. Our approach obtains similar results to specific approaches in terms of the number of solved problems, with reasonable requirements in computational resources.

The Capacity Assignment (CA) problem focuses on finding the best possible set of capacities for the links that satisfies the traffic requirements in a prioritized network while minimizing the cost. Most approaches consider a single class of packets flowing through the network, but in reality, different classes of packets with different packet lengths and priorities are transmitted over the networks. In this paper we assume that the traffic consists of different classes of packets with different average packet lengths and priorities. Marayuma and Tang [7] proposed a single algorithm composed of several elementary heuristic procedures. Levi and Ersoy [6] introduced a simulated annealing approach which produced substantially better results. A new method that uses continuous learning automata was introduced in a previous paper [12]. In this paper we introduce a new method which uses discretized learning automata to solve the problem. Indeed, to the best of our knowledge, this is the fastest and most accurate scheme currently available.

We present a new version of our robot manipulator planning toolbox, OXSIM. OXSIM is designed to greatly simplify the building of planners by providing core competence in three-dimensional geometry. This is done by the provision of efficient routines for computing the distance between parts of the robot and its environment. A new version of OXSIM, written in C++, provides an object-oriented interface to the basic system, which will increase its ease of use.

We present a parallel path planning method that is able to automatically handle multiple goal configurations as input. There are two basic approaches, goal switching and bi-directional search, which are combined in the end. Goal switching dynamically selects a favourite goal depending on some distance function. The bi-directional search supports the backward search direction from the goal to the start configuration, which is probably faster. The multi-directional search with goal switching combines the advantages of goal switching and bi-directional search. Altogether, the planning system is enabled to select one of the preferable goal configuration by itself. All concepts are experimentally validated for a set of benchmark problems consisting of an industrial robot arm with six degrees of freedom in a 3D environment.

We present a novel method for robot motion planning that constructs a network of collision free paths using a randomized search over a potential field in Configuration Space. Our method finds local minima and then connects them to form a graph, which we call a roadmap. We use a gradient search scheme to find the local minima very efficiently and accurately. To find a path between two configurations, it is then a simple matter to connect given start and goal configurations to the roadmap and to use a standard graph search algorithm to search the roadmap. The construction of the roadmap can be done in parallel with very little communication.

Rather than presenting a specific technique to plan free motions, the main purpose of this paper is discussing the potential of possible approaches from different viewpoints. The focus is on planning simple motions on the basis of a fine grain description of the workspace. We consider the problem of planning translations of a convex polygon in a cluttered polygonal environment, i.e., in the presence of several convex bodies with several sides, as a toy example to address a number of questions: limits of some popular approaches, development of more refined—but practical—techniques, comparison between algorithmic and intuitive motion planning, use of dynamic techniques, potential of parallelization. Most of the ongoing considerations will take the results of a few numerical experiments as their starting point.

A method based on the union of an Evolutionary Algorithm (EA) and a local search algorithm for obtaining coordinated motion plans of two manipulator robots is presented. A Decoupled Planning Approach has been used. For this purpose, the problem has been decomposed into two subproblems: path planning, where a collision-free path is found for each robot independently of the other, only considering fixed obstacles; and trajectory planning, where the paths are timed and synchronized in order to avoid collision with the other robot. This paper focuses on the second problem. A method is presented to minimize the total motion time of two manipulators along their paths, avoiding collision regardless of the accuracy of the dynamic model used. A hybrid technique with EA and local search methods has been implemented.

This paper proposes a solution by genetic algorithms to the problem of planning a robust and suboptimal trajectory in the velocity space of a mobile robot. Robust trajectories are obtained introducing cumulative noise in the evaluation of the fitness function and introducing modifications in the genetic algorithm to taking into account this new feature. Results are presented that show the performance of the algorithm in different environments and the influence of the noise in the planned trajectories.

We present a practical method for coordinating two robots that have to move in close proximity. Few results have been reported in the literature for practical situations under time constraints. In addition, no previous approaches deal with the problem of coordinated motion when the goal configurations are time dependent. We present a solution to a particular case of the multimovers problem: the coordination of two robots when grasping objects of unknown shape and position from a conveyor belt. The approach take advantage of a collision detection algorithm that uses a hierarchy of spheres to represent the robots. Experimental results yielding feasible computation times are presented.

This paper presents a self-organizing approach for mobile robot path planning problems in dynamic environments by using case-based reasoning together with a more conventional method of grid-map based path planning. The map-based path planner is used to suggest new innovative solutions for a particular path planning problem. The case-base is used to store the paths and evaluate their traversability. While planning the route those paths are preferred which, according to former experience, are least risky. As the environment changes, the exploration as well as the evaluation of the paths will allow the system to self-organize by forming a set of low-risk paths that are safest to follow. The experiments in a simulated environment show that the robot is able to adapt in a dynamic environment and learns to use the least risky paths.

Configuration is the process of composing a system from a set of components such that the system fulfills a set of desired demands. The configuration process relies on a particular component model, which is a useful abstraction of the domain and the technical system to be composed.In this place we deal with configuration problems where the components involved are characterized by simplified functional dependencies, so-called resource-based descriptions. On the one hand, the resource-based component model provides for powerful and user-friendly mechanisms to formulate configuration tasks. On the other hand, the solution of resource-based configuration problems is NP-complete, which means that no efficient algorithms exist to solve a generic instance of that problem.In practice, given a concrete resource-based component model, the search for an optimum configuration can be realized efficiently by means of heuristics that have been developed by domain experts. The paper in hand picks up that observation: It presents a method to automatically generate heuristics that guide the search when solving complex resource-based configuration problems.

We compare the use of stochastic dynamic programming (SDP), Neural Networks and a simple approximation rule for calculating the real option value of a flexible production system. While SDP yields the best solution to the problem, it is computationally prohibitive for larger settings. We test two approximations of the value function and show that the results are comparable to those obtained via SDP. These methods have the advantage of a high computational performance and of no restrictions on the type of process used. Our approach is not only useful for supporting large investment decisions, but it can also be applied in the case of routine decisions like the determination of the production program when stochastic profit margins occur.

We consider the problem of designing an optimal control for linear discrete-time systems assuming that total cost for the control efforts is limited and cost function is periodic in behaviour. (seasonal, for example) This Model was developed as a result of analysis of real data of the project “Modelling River Murray Estuary” from the Environmental Modelling Research Group, the University of South Australia.Nonlinear control systems are of a great significance in the field of control engineering since most practical dynamic systems are nonlinear. Using arbitrary control strategy as an initial we can compute coefficients in nonlinear system as a function of corresponding output variables. (previous output variables) As a result we shall transform nonlinear system into linear system with known optimal solution. (new output variables) Repeating this procedure again and again we shall generate sequence of control strategies. Optimal control strategy for given nonlinear system may be obtained as a limit of this sequence. This fact has been demonstrated by the particular example relating to the above environmental research Project.

In this paper we present a scheduler suitable to be applied to a particular class of dynamic systems, which main characteristics are the lack of actual data during long time periods and the unreliability on the available data. The management of these systems requires the integration of simulation techniques, temporal reasoning, soft real-time and mechanisms of reason maintenance. To deal with all of these qualities it is showed a centralized hierarchical task scheduler, which main operation characteristics are event oriented and hierarchical task oriented. We apply this scheduler to a deep knowledge expert system developed for monitoring and helping to the decision taking in an urban traffic control system.

To solve the difficulties of numerical integral that exist in global optimization of redundant manipulators, this paper discusses how to build up dynamic equation, the inner relation between constraints and unconstraint optimum control problems, then a numerical method solving optimum control problem is deeply analyzed. A directiteration method in solving normal equations by means of bidirectional asynchronous integral has been successfully exploited, so that it can efficiently overcome the difficulty in solving two-point boundary value problems resulting from inverse stability between state equation and co-state equation.

The concept of soft computing (SC) and Artificial Intelligence methods (AI) for design of ship-board intelligence system (IS) of safety navigation and submarine dynamic object control are discussed. The syntheses of the underwater vehicle control rules is considered as practical application on the base of fuzzy logic elements and neural networks (NN) logic base. The genetic algorithm is used for forming NN training succession. Control law topology corresponds to multilayer NN.

A diagnostic structure based on modular expert system architecture is presented in this paper. An illustrative example shows how expert knowledge is used to deduce fault situations. In this example, qualitative representation of numerical variables of process is used to interface expert knowledge and process variables. The goal is to use this new information to deduce abnormal situations that can not be differentiate from the normal evolution of process dynamics using only data available from process. The evolution of process variables and abstracted information are used as inputs for the reasoning modules.

This contribution addresses the following supervisory problem: a continuous plant (P) is to be supervised via symbolic (or quantised) actions. These symbolic actions suggest the set points to the lower level control loops. The system dynamic is analysed on supervisory level (K) by a qualitative approach. The relationships between variables and the steady-state references are known. These problems are especially common in chemical process control. The supervisor handles start-up and shut-down procedures and takes appropriate action to solve the sequential or parallel tasks of a basic procedure. The object of this paper is to introduce an approach to solving the problem of how to derive a set of rules from a physical process.

The solutions to supervise start-up and shut-down operations in close loop are suitable for large industrial systems. Similarly, the batch and semi-continuous processes in order to maintain the operation in a dynamic mode. This paper considers the qualitative event-based expert supervision approach of these problems of a distillation column. The development of a general supervision in this work is based on an events generator and on a corrective actions generator. The qualitative symbols are based on fuzzy sets. In particular there are mechanisms to process the changes in the system variables from qualitative symbols.

Supervision of power distribution systems is a major concern for electricity distributors. It consists in analysing alarms emitted by the devices, detecting permanent faults, locating these faults and reconfiguring the network to maintain the optimal quality of service. EDF is developing the AUSTRAL platform which provide distribution control centers with a set of advanced real-time functions for alarm processing, diagnosis and power supply restoration. After an overview of the general architecture of AUSTRAL, this paper focuses on the alarm processing and restoration functions, which both use model-based approaches.

This paper deals with a behavioral interpretation of physical systems. It describes how we can interpret the system current state using a multimodeling principle. The system representation proposed by the multimodeling approach is based fundamentally on the qualitative and functional knowledge generated according to a bond graph description.

An intelligent agent is “anything that can be viewed as perceiving its environment through sensors and acting upon its environment through effectors”. An agent performs its work by using rules (or knowledge) that maps the information its sensors give it to actions it is supposed to perform. In order to assist system administrators with some very important yet routine tasks, an intelligent agent was developed. The agent monitors certain system parameters and situations, decides if corrective action is warranted, and carries out the corrective action. Four of the most important UNIX system administrator tasks are the focus of this intelligent agent. These tasks are: detecting inactive accounts, changing the priority of CPU-intensive processes when the system load is high, deleting old files (such as core files or old backup files), and trimming the wtmp and wtmpx files.

We present a real-time knowledge based [8] supervision support system in the coal washing domain. The Ash Control Model (AshMod) assists operators in maximising clean coal yield while keeping ash (impurity) content within acceptable limits. AshMod assists the operator in plant situation assessment, fault diagnosis, and performance optimisation.In this paper, we describe the optimisation task, which employs a hybrid artificial intelligence and operations research approach. The process is modeled through a set of extended states associated with the entire process and with individual components (circuits) within the plant. The process is continuously monitored to assess the process state, which dynamically influences the planning and scheduling of a sequence of optimisation steps.The supervision support system captures domain knowledge through multiview knowledge models [7] that capture purpose, function, structure, behaviour and heuristics. The supervision support system is currently undergoing online validation at the B&C Coal Washing Plants operated by the Broken Hill Proprietary Limited (BHP) at Port Kembla, Australia [1].

Correct climate control improves the quality of productions in greenhouses. Those control techniques that do not take into account the non-linear and multivariable features of the climate in the greenhouse, cannot achieve good performance (set-points will not be accomplished). This paper presents a Predictive Contro based technique using a mathematical model of the climate behaviour and Simulated Annealing as optimizer. Results show that this technique can be useful when dealing with non-linear and multivariable plants, even if constraints in the control actions are considered.

This paper presents an approach to free space and obstacle detection based on stereo disparity, with applications in mobile robotics. Two maps of an area of interest on the ground plane are computed from the left and right views. Obstacles rising up from the ground occlude different areas in both maps. An obstacle map is computed as the binarized difference between the two maps. Morphological filtering is used to detect the free space from the obstacle map, while obstacles are detected by looking for instances of a suitable obstacle model. Results of the approach in road scenes are presented. Stereo images are taken from a car, with the objective of detecting other cars and free space in cluttered scenes for driving assistance.

In this paper a novel method of reference places' detection to build topological models is described, as well as an algorithm for route planning based on Fuzzy Petri Nets. The proposed method for the detection of reference places does not employ sensory information but the information provided by the robot's control subsystem. The reference places and the navigation strategies between places are used for constructing the model's environment with a Fuzzy Petri Net. The route planning algorithm propagates over the net the certainty value of places and transitions. After finishing the propagation the transitions and the places store the information needed to make decisions about the navigation strategies of the robot route.

In this work we describe the design of a control approach in which, by way of supervised reinforcement learning, the learning potential is combined with the previous knowledge of the task in question, obtaining as a result rapid convergence to the desired behaviour as well as an increase in the stability of the process. We have tested the application of our approach in the design of a basic behaviour pattern in mobile robotics, such as that of wall following. We have carried out several experiments obtaining goods results which confirm the utility and advantages derived from the use of our approach.

Mobile robots usually employ dedicated networks to share data and control in an open environment. These networks, called adhoc networks, are multi-hop wireless networks where the mobile devices communicate using a shared, low power, low bandwidth channel. Since the classical routing algorithms of the wired networks cannot be used, new communication and routing protocols are being developed. In this paper we present a new communication protocol that solves data transfer problems, like packet routing, in an ad-hoc network used to control a cloud of robots. This protocol results to be efficient in a situation where a dedicated node, called the coordinator, controls the activity of all the robots, therefore being one of the end-points of any data interchange.

The objective of this work is to study neural control architectures for autonomous robots that explicitly handle time in tasks that require reasoning with the temporal component. The controllers are generated and trained through the methodology of evolutionary robotics. In this study, the reasoning processes are circumscribed to data provided by light sensors, as a first step in the process of evaluating the requirements of control structures that can be extended to the processing of visual information provided by cameras.

This paper addresses adaptive, on-line path planning of an autonomous underwater vehicle and presents a GA-based method for it. It is an important module of SAUVIM (Semi-Autonomous Underwater Vehicle for Intervention Missions) which is being developed at the University of Hawaii and will be capable of exploring the ocean at up to 6,000 m depth. In SAUVIM, a genetic algorithm (GA) is employed in order to integrate on-line path planning with off-line planning and make path planning adaptive. We first discuss how sensory information is incorporated into pre-loaded mapping data of the ocean floor. Then, we present a method for updating a path in real time while the vehicle is moving. A prototype of the adaptive, on-line path planning module is also presented.

Current robotics is perhaps the most complete paradigm of applied Artificial Intelligence, since it includes generic tasks involving pluri-sensorial integration and internal representation, as well as motor planning and control. In this paper we revise the architecture proposed by Craik and McCulloch and the concept of environment model introduced by K. Craik. Based on this architecture, which links the description in terms of properties with the selection of a mode of action, we study a simple example application in which an incremental procedure is proposed for the construction and use of a model of a structured medium (the interior of a building) using a graph. The type of graph used to store the descriptions of objects and the relations between them is inspired by the work of Hillier and Hanson on the analysis of interiors. The connections between the elements of the environment (graph nodes) are generated in such a way as to facilitate their efficient use for the selection of the most pertinent mode of action at any given moment. The derivation of the graph is carried out autonomously. In the development of this work, we have avoided as far as possible the use of anthropomorphic terms with no causal connection to the symbol level. Posed in this way, the problem of the representation and use of an environment model by a robot reduces to the use of models of generic tasks and methods at the knowledge level together with graphs and finite state machines at the formal level.

Although telerobotic systems are becoming more complex, there are few actions they can perform on their own and, moreover, knowledge about the tasks they are being used for often relies only on their operator. In this paper, we present the design of a telerobotic system that features learning capabilities, can accept commands given in natural language and provides control of a generic vision-guided robot. This system is composed of a set of functional blocks that communicate among them according to the CORBA standard over an Internet-based network. Knowledge is originated from interaction with users and is stored in a shared database. The user interface translates user commands into a set of predefined commands that the robot controller can understand, making it unnecessary for users to know which particular robot they are working with. We describe the design of this system and focus on two of its main components: the user interface and the robot controller.

Stereoscopic vision is an appropiate tool for building maps of the environment of a robot. When matching regions of the images, segmentation errors should be avoided. In this paper an algorithm to deal with errors in region matching is proposed, and the results in the presence of noise are analyzed. The selection of an appropiate similarity criterion to create the initial nodes in the graph-based matching process is very important for reducing the time of computation considerably. The experimental results show that the method is robust in the presence of noise.

Functional reasoning or the Takagi-Sugeno-Kang model is a fuzzy reasoning method aiming at numerical accuracy and has found wide use in fuzzy modeling. In this method, each rule consists of a fuzzy implication and a functional consequence part. In this work, a new, online identification method for such a system is presented, for supervised learning tasks. Structure identification is executed by a fuzzy ART learning module, following the procedure of splitting fuzzy rules that tend to produce high output error. Fuzzy rules are also added wherever the error exceeds a threshold. All parameters are fine tuned by the δ rule, a basic learning technique in neural networks. Computer simulation exhibits the potentials of this approach, which is tested with well known benchmarks, yielding excellent results.

In this paper, Adaptive Neuro Fuzzy Inference System (ANFIS) is used for the controlling of a commercial robot manipulator. A Microbot [1] with three degrees of freedom is utilized to evaluate the proposed methodology. A decentralized ANFIS controller is used for each joint, with a Fuzzy Associative Memories (FAM) performing the inverse kinematics mapping in a supervisory mode. The individual fuzzy controller for each joint generates the required control signal to a DC servo motor to move the associated link to the new position. The simulation experiments indeed demonstrate the effectiveness of the proposed method.

The users of libraries of object-oriented software components face with both a terminological and a cognitive gap. Usually, library users do not understand the vocabulary used in the library documentation, and they do not know some of the concepts involved in the solutions implemented by the components. In order to close both gaps, we define a language to construct functional descriptions of the components, allowing the users to specify their needs. From this language, we implement the mappings among functional descriptions and components as a case base collecting “interesting experiences” in the usage of the library. We have used a knowledge representation system based on description logics to build the description language and to implement the case-based reasoning processes. We have applied this approach to support software reuse in the class library of VisualWorks, a Smalltalk programming environment.

In this paper we are interested in knowledge-based techniques (called program supervision) for managing the re-use of a modular set of programs. The focus of this paper is to analyse which re-use problems program supervision techniques can solve. First a general definition for program supervision and a knowledge representation model are proposed. Then an analysis is presented in terms of the structure of the programs to re-use and in terms of the effort for building a program supervision knowledge base. This paper concludes with what program supervision can do for program re-use from the points of view of the code developers, the experts, and the end-users.

One important approach to enhancing software reuse is through the creation of large-scale software libraries. By modularizing functionality, many complex specialized applications can be built up from smaller reusable general purpose libraries. Consequently, many large software libraries have been formed for applications such as image processing and data analysis. However, knowing the requirements and formats of each of these routines requires considerable expertise — thus limiting the usage of these libraries by novices.This paper describes an approach to enable novices to use complex software libraries. In this approach, the interactions between and requirements of the software modules are represented in a declarative language based on Artificial Intelligence (AI) Planning techniques. The user is then able to specify their goals in terms of this language - designating what they want done, not how to do it. The AI planning system then uses this model of the available subroutines to compose a domain specific script to fulfill the user request. Specifically, we describe three such systems developed by the Artificial Inteligence Group of the Jet Propulsion Laboratory.The Multimission VICAR Planner (MVP) has been deployed for 2 years and supports image processing for science product generation for the Galileo mission. MVP has reduced time to fill certain classes of requests from 4 hours to 15 minutes. The Automated SAR Image Processing system (ASIP) is currently in use by the Dept. of Geology at ASU to support aeolian science analysis of synthetic aperture radar images. ASIP reduces the number of manual inputs in science product generation by 10-fold. Finally, the DPLAN system reconfigures software modules which control complex antenna hardware to configure antennas to support a wide range of tracks for NASA's Deep Space Network of communications and radio science antennas.

In this paper, we examine some aspects related to the knowledge-based control of vision systems. We distinguish between the specialist and user modes of control, based on the granularity at which the control takes place. We illustrate the user mode of control on an application developed using the PEGASE [5] framework developed at INRIA Sophia-Antipolis.

SOLUTION is a knowledge based system, which can be used to automatically configure and adapt the low level part of image processing systems with respect to different tasks and input images. The task specification contains a characterization of the properties of the class of input images to be processed, a description of the relevant properties of the output image to be expected, requests about some general properties of the algorithms to be used, and a test image. In the configuration phase appropriate operators are selected and processing paths are assembled. In a subsequent adaptation phase the free parameters of the selected processing paths are adapted such that the specified properties of the output image are approximated as close as possible. All task specifications including the specification of the requested image properties are given in natural spoken terms like the Thickness or Parallelism of contours. The adaptation is rule based and the knowledge needed therefore can be learned automatically using a combination of different learning paradigms. This paper describes the adaptation and the learning part of SOLUTION.

A new machine learning system, INNER, is presented in this paper. The system starts out from a collection of training examples; some of them are inflated generalizing their description so as to obtain a first draft of classification rules. An optimization stage, borrowed from our previous system, Fan, is then applied to return the final set of rules. The main goal of Inner, besides its high level of accuracy, is its ability for self-maintenance. To close the paper, we present a number of different experiments carried` out with INNER to illustrate how good the performance and stability of the system is.

In this paper we describe TIM (Total Induction Method), a framework that empowers inductive learning in real domains by the construction of new higher level features based on the relations between the descriptors of the initial training set. A new method, named FDD, for discovering functional dependencies within the data is outlined, and details regarding its relevance for constructive learning are provided. Two examples of their application in real - world domains are given.

This paper proposes an algorithm to design a tree-like classifier whose result is equivalent to that achieved by the classical Nearest Neighbour rule. The procedure consists of a particular decomposition of a d-dimensional feature space into a set of convex regions with prototypes from just one class. Some experimental results over synthetic and real databases are provided in order to illustrate the applicability of the method.

This paper analyses the viability in use a self-organizing metaheuristic known as Elastic Network to solve the Euclidean Traveling Salesman Problem (ETSP). We propose a new hybrid approach based on Elastic Networks associated with a convex hull algorithm to solve the ETSP. In order to analyze our hybrid algorithm, one of the best Elastic Networks versions was also implemented. Experimental results show the quality of our algorithm is good.

During the last few years, we have developed an expect system, called PORSEL (PORtfolio SELection system), which uses a small set of rules to select stocks. This paper improves the PORSEL by incorporating several new features and interfaces. The new PORSEL now consists of three components: the information center, the fuzzy stock selector, and the portfolio constructor. The purpose of the information center is to provide representation of several technical indicators such as candlestick charts, moving average of closing prices, and price trends. The fuzzy stock selector evaluates the listed stocks and then assigns a composite score for each stock. The portfolio constructor generates the optimal portfolios for the selected stocks. The new PORSEL also includes a user-friendly interface for adding and deleting rules during the run time. The results of simulation show that our new version of PORSEL outperformed the market almost every year during the testing period.

This paper presents a survey of different adaptation techniques and the used knowledge during adaptation. A process model of CBR and the used knowledge according to the different knowledge containers is introduced. The current models of adaptation are described and illustrated in an example domain.

As case-based reasoning systems are deployed in real-world situations the issue of case maintenance becomes more and more critical. Uncontrolled case-base growth can cause serious performance problems as retrieval efficiency degrades and incorrect or inconsistent cases become increasingly difficult to detect. This paper surveys recent progress in the area of knowledge maintenance, and propose a novel, competence-based maintenance policy for case-based reasoning systems.

There is considerable enthusiasm about Case-Based Reasoning as a means of developing knowledge-based systems. There are two broad reasons for this enthusiasm. First, it is evident that much of human expert competence is experience based and it makes sense to adopt a reuse-based methodology for developing knowledge based systems. The other reason is the expectation that using Case-Based Reasoning to develop knowledge based systems will involve less knowledge engineering than alternative 'first-principles' based approaches. In this paper I explore the veracity of this assertion and outline the types of situation in which it will be true. CBR is perceived to have this knowledge engineering advantage because it allows the development of knowledge based systems in weak theory domains. If CBR can work without formalising a domain theory then there is a question about the quality of solutions produced by case-based systems. This is the other issue discussed in this paper and situations where CBR will and will not produce good quality solutions are outlined.

This paper asks whether case-based reasoning is an AI technology like rule-based reasoning, neural networks or genetic algorithms or whether it is better described as a methodology for problem solving, that may use any appropriate technology. By describing four applications of CBR, that variously use: nearest neighbour, induction, fuzzy logic and SQL, the author shows that CBR is a methodology and not a technology.

A case base is a repository of past experiences that can be used for problem solving. Given a new problem, expressed in the form of a query, the case base is browsed in search of “similar” or “relevant” cases. One way to perform this search involves the iterative evaluation of a series of queries against the case base, where each query in the series is obtained by restricting or relaxing the preceding query.The paper considers alternative approaches for implementing iterative browsing in case-based reasoning systems, including a naive algorithm, which evaluates each query independent of earlier evaluations, and an incremental algorithm, which reuses the results of past query evaluations to minimize the computation required for subsequent queries. In particular, the paper proposes an efficient algorithm for case base browsing and retrieval using database techniques for view maintenance. In addition, the paper evaluates the performance of the proposed algorithm with respect to alternative approaches considering two perspectives: (i) experimental efficiency evaluation using diverse application domains, and (ii) scalability evaluation using the performance model of the proposed system.

This paper proposes an efficient case based method of constructing a coal molecular structure model from many pieces of threedimensional block data. In this method, the block data, which is a partial structure that consists of a few aromatic fragments, are optimized in terms of steric energy and are stored in the case base in advance. This method consists of the following two processes: (1) retrieving suitable three-dimensional block data from the case base, and (2) combining them involving less stress. In order to effectively retrieve well-matched block data, the three-dimensional similarity between the blocks is evaluated. We found that this method can derive molecular structures more scientifically and in more quickly than the hand-made structures.

A problem faced by many constructive neural networks using a cascade architecture is the large network depth. This results in large fan-in and propagation delays, problems especially relevant for VLSI implementation of these networks. This work explores the effect of limiting the depth of the cascades created by CasPer, a constructive cascade algorithm. Instead of a single cascade of hidden neurons, a series of cascade towers are built. Each cascade tower can be viewed as a single Higher Order Neuron (HON). The optimal complexity of the HON required for a given problem is difficult to estimate, and is a form of the bias-variance dilemma. This problem is overcome via the construction of HONs with increasing complexity. It is shown that by constructing HONs in this manner the chance of overfitting is reduced, especially with noisy data.

Artificial Neural Networks (ANNS) are used successfully in industry and commerce. This is not surprising since neural networks are especially competitive for complex tasks for which insufficient domain-specific knowledge is available. However, interpretation of models induced by ANNS is often extremely difficult. BP-SOM is an relatively novel neural network architecture and learning algorithm which offers possibilities to overcome this limitation. BP-SOM is a combination of a multi-layered feed-forward network (MFN) trained with the back-propagation learning rule (BP), and Kohonen's self-organizing maps (sorts). In earlier reports, it has been shown that BP-SOM improved the generalization performance as compared to that of BP, while at the same time it decreased the number of necessary hidden units without loss of generalization performance. In this paper we demonstrate that BP-SOM trained networks results in uniform and clustered hidden layer representations appropriate for interpretation of the networks functionality.

The problem of weight initialization in multilayer perceptron networks is considered. A computationally simple weight initialization method based on the usage of reference patterns is investigated in channel equalization application. On one hand, the proposed method aims to set the initial weight values to be such that inputs to network nodes are within the active region. On ] the other hand, the goal is to distribute the discriminant functions formed by the hidden units evenly into the input space area where training data is located. The proposed weight initialization is tested in the channel equalization application where several alternatives for obtaining suitable reference patterns are investigated. A comparison with the conventional random initialization shows that significant improvement in convergence can be achieved with the proposed method. In addition, the computational cost of the initialization was found to be negligible compared with the cost of training.

As yield and productivity are increasingly competing in importance with technology in integrated circuit manufacturing, semiconductor industry can benefit from advances on artificial intelligence. This paper shows a fault diagnosis system based on autoassociative neural networks, a little exploited processing architecture in industrial applications. The system integrates three autoassociative algorithms and it selects the most suitable in each case. It optimizes the processing time while guarantees an accurate diagnosis. The feasibility of the solution is justified and comparative results are presented and discussed.

A method for training of an ML network for classification has been proposed by us in [3,4]. It searches for the non-linear discriminant functions corresponding to several small local minima of the objective function. This paper presents a comparative study of our method and conventional training with random initialization of the weights. Experiments with a synthetic data set and the data set of an OCR problem are discussed. The results obtained confirm the efficacy of our method which finds solutions with lower misclassification errors than does conventional training.

A new architecture for unsupervised learning is proposed. The topology, activation rules, and training algorithm are presented and a specific training base is used to prove the advantages of this type of network. The training patterns are from chess playing, but there are several other applications for this kind of system, and a specific one is proposed without going into details. Experimental results emphasize the performances of the network training.

Reinforcement learning, in general, has not been totally successful at solving complex real-world problems which can be described by nonlinear functions. However, temporal difference learning is a type of reinforcement learning algorithm that has been researched and applied to various prediction problems with promising results. This paper discusses the application of temporal-difference-learning in training a neural network to play a scaled-down version of Chinese Chess. Preliminary results show that this technique is favorable for producing desired results. In test cases where minimal factors of the game are presented, the network responds favorably. However, when introducing more complexity, the network does not function as well, but generally produces reasonable results. These results indicate that temporal difference learning has the potential to solve real-world problems of equal or greater complexity. Continuing research will most likely lead to more responsive and accurate systems in the future.

The Cardiovascular Diseases (CVD) have a multifactorial origin. They are caused by mixed effects of some gene and environmental factors. Therefore, gene identification strategies needed to discover involved gene regions are more complex than strategies applied on monogenic diseases. Nowadays, the first ones lack satisfactory levels of performance and efficiency.We present a “in progress” work with the aim to design computational strategies to perform gene screenings to identify new genomic regions associated to atherogenic CVD process. In this work we apply formal methodology and techniques related to logic programming, object-oriented development architectures and combinatorial aspects of genetic model. We have implemented an alpha version of the system which has the main functions designed. And we are actually designing a complete set of data test based on the Framingham Heart Study to compare with classical linkage tools.

This paper, presents a system, SISTER, that is suitable for storing and retrieve of large collections of images allowing the user to formulate queries for different image categories on the basis of color information and specific attributes of the image category; this is possible, because SISTER can be easily adapted to support new image categories specializing the acquisition and the retrieval subsystems. SISTER is composed of three parts: i) an image acquisition subsystem automatically extracting the attributes from images, ii) a database management subsystem maintaining the descriptions of images, and iii) a retrieval subsystem allowing the user to retrieve images through a user-friendly graphical interface. The acquisition subsystem extracts image attributes combining image processing and inductive classification modules. Classification modules are useful because they may allow the extraction of attributes that cannot be extracted through image processing techniques and because they may allow the reduction of the error percentage made by the image processing modules in the acquisition of image attributes.

Severely handicapped people are those who lack at least 70% of their physical and/or psychological functionalities (visual, hearing, physical and cognitive/language impairments). One of the main problems is the communication between them and with non handicapped people. However, some individuals, assisted and helped by their trainers, can reach normal levels of educationWe propose a multimedia system (COMBLISS: Communication Bliss System) based on man-machine interface and object-oriented databases which aims to enhance handicapped people skills related to communicating with other people.The conceptual architecture of our multimedia system is composed by three levels. Bottom: the conceptual schemata of databases with Bliss symbols and user context profile. Middle: a set of conceptual objects to model the real world. Upper: multimedia presentations of conceptual objects as representation objects.Prototype results are slightly poor because it is not possible to show effectively more than 40 Bliss symbols per screen shot. But users must be trained before work with efficiency.

We have proposed Meta-Museum as a new knowledge sharing environment where experts and novices can communicate with each other with agent support. Museum exhibitions are thought to be well organized representations of the expert knowledge of curators, but they are just one example of structures of knowledge among many possibilities, given to museum visitors in a one-sided way. Therefore, traditional museum exhibitions can hardly meet the vast requirements of general visitors who possess a variety of interests. In this paper, we propose agents to mediate between curators and visitors, so that both sides can convey their interests and knowledge to one another and gain a better understanding. These mediating agents visualize the semantic relations of displays as a two-dimensional spatial structure based on the viewpoints of the curators and visitors separately, and then together. The structures reflect the interests of the visitors, while maintaining the knowledge of the curators.

In this paper we introduce a probabilistic approach to support visual supervision and gesture recognition. Task knowledge is both of geometric and visual nature and it is encoded in parametric eigenspaces. Learning processes for compute modal subspaces (eigenspaces) are the core of tracking and recognition of gestures and tasks. We describe the overall architecture of the system and detail learning processes and gesture design. Finally we show experimental results of tracking and recognition in block-world like assembling tasks and in general human gestures.

We present an approach to create structured and shared workspaces supplying functionality to carry out collaborative activities for a range of learning tasks. A workspace is defined in terms of three interrelated components: group, task and colaboration; each one modelling respectively the relevant features for designing the structure and support for a collaborative learning activity. The system has been used to create a number of applications addressing different educational goals and target groups.

The application of computers for the planning and operation of water resource systems is a rapidly advancing field of research. In recent years, decision support system (DSS) has gained much attention in civil engineering, in which the output can be displayed in high quality and easy to be understood. In this study, the decision support system for integrated water management, CTIWM, is developed with particular reference to Chikugo River basin, a multipurpose multireservoir system in Japan. It uses a module library that contains compatible modules for simulating a variety of water and physio-chemical processes. Different kinds of numerical models may be invoked through user interface menu, which facilitates communications between users and models in a friendly way. It demonstrates that the integration of DSS technique, simulation and optimization models is an efficient way for water resources management.

A decision model for railway interlocking systems (independent of the topology of the station) is presented. The safety of a situation is decided by checking (using Gröbner Bases) whether or not a certain ideal of a polynomial ring has degenerated into the whole polynomial ring. This ideal somehow translates the oriented graph associated to the situation of trains, switches and signals (or semaphores).If a section is accessible by a train located in another given section can also be checked by testing an ideal membership. The fact that trains could occupy more than one section does not affect the model.The authors have developed a method for dealing with verification and knowledge extraction in Expert Systems [3]. Such a method, altered, is reused in the decision support methods used both in this article and in an application to appropriateness criteria in Medicine [6].

Isokinetics systems are now a leading technology for assessing muscle strength and diagnosing muscle injuries. These systems are very expensive, for which reason they should be put to the best possible use. However, the computer interfaces that come with isokinetics systems are extremely poor and do not provide for the system to be exploited to its full potential. This paper presents the project 14 (Interface for Intelligent Interpretation of Isokinetic Data) and two computer systems obtained in the project: ISODEPOR and ISOCIN Both applications provide simple and effective interaction with the LIDO Isokinetics Machine, that produces a huge amount of strength data in each isokinetics test. These data are interpreted and presented to the user, who interacts with the information by means of a powerful graphic display system. Additionally, the applications have been equipped with a series of intelligent strength data analysis functions that implement expertise.

HADES, a knowledge based system for interpretation and fusion of formatted military messages in the context of land battle will be presented. Data and information fusion based on military experience and heuristic rules has been implemented in an automated system. The scope of the project and main research problems are defined and a simulation model is presented. An overview of the architecture of the system and the general processing flow is given. The main functionalities and fusion methods are outlined. Finally some comments on the evolution of the system and the results are made.

The paper presents the research framework for the design of a special software environment to support visual knowledge base design and specification. Flexible user centred graphical interface is described. The approach is aimed at three interrelated topics: knowledge specification, visual structuring and hypertext design. The proposed ideas and methods may be applied to those systems where structured analysis of data and knowledge is of special interest, such as intelligent tutoring systems, expert systems, decision support, etc.

A specific tool for evaluating and tutoring students has been developed. The application results help the students to review those subjects on which their larnowlegde level is not adequate and assist tutors to direct their courses to the corresponding level of a student or a group of students. The tool implementation is made possible by the development of a questions data base in the specific subject to be evaluated.

The difficulty of designing and developing more useable and cost-effective intelligent tutoring systems (ITSs) has caused the realization of some new approaches in that field, the realization of intelligent tutoring shells - authoring tools. Our starting point and perspective on developing ITSs shell is motivated by issues of pragmatics and usability. The paper describes pedagogical aspects and the control knowledge of the component based model of intelligent tutoring systems, called GET-BITS. The focus is on class hierarchies and design of classes for representing the pedagogical and the control knowledge in the GET-BITS model_ Distinction has been made between different modes of operation of the pedagogical module in an intelligent tutoring system (teaching, examination, and consulting), resulting in the necessity for the class hierarchies to reflect specific pedagogical knowledge structure.

This paper presents a self-improving reactive control system for autonomous agents. The design process consists of three main parts: first, building a self-organizing map and integrating the available knowledge about the system into the neural control structure, second, improving the performance of the agent with regard to the individual goals separately, and third, combining the obtained results to get an optimal overall behavior of the system. In this paper the emphasis is put on the second part. Improvement consists of identifying the dynamics of the environment using exploration and determining an optimal behavior selection policy using techniques of dynamic programming. We show the effectiveness of the improvement method and evaluate it through several simulation studies.

Management of complexity, changes and disturbances is one of the key issues of production today. Distributed, agent-based structures represent viable alternatives to hierarchical systems provided with reactive / proactive capabilities. In the paper approaches to distributed manufacturing architectures are surveyed, their fundamental features are highlighted. An object oriented simulation framework for development and evaluation of multi-agent manufacturing architectures is introduced together with an approach to agent-based scheduling. Further research issues are highlighted.

The application of pattern recognition (PR) techniques, artificial neural networks (ANNs), and nowadays hybrid artificial intelligence (Al) techniques in manufacturing can be regarded as consecutive elements of a process started two decades ago. The fundamental aim of the paper is to outline the importance of soft computing and hybrid AI techniques in manufacturing by introducing a genetic algorithm (GA) based dynamic job shop scheduler and the integrated use of neural, fuzzy and GA techniques for modeling, control and monitoring purposes.

In the literature there exist several soft computing methods for building predictive models: neural network models, fuzzy models and probabilistic approaches. In this paper we are interested in the question which one of these approaches is likely to give best performance in practice. We study this problem empirically by selecting a set of typical models from the different model families, and by experimentally evaluating their predictive performance. For the evaluation, we use two real-world manufacturing datasets from a production plant of electrical machines. The models considered here include fuzzy rulebases, various neural network models and probabilistic finite mixtures. Our investigation indicates that all the methods can produce predictors that are accurate enough for practical purposes. Moreover, the results show that adding expert knowledge leads to improved predictive performance in the domain where such knowledge was available. In the domain where no expert knowledge was available, the probabilistic approach produced the best results.

This paper presents a simulation platform that has been realised in Java 1.1 for the study of behaviour and evolutionary processes in non-linear systems. To support the modelling of such systems, we propose the use of agent technology as high level tool to design applications. The framework enables to study emergence by exploiting distributing computing as key issues of the system behaviour. Applications developed with the platform are then simulated to adequately capture any behaviour likely to be observed, to exhibit self organised structures, and to emphasise complex processes, which are brought in action. This approach then allows the study of macroscopic collections endowed with the potential to evolve during time.

This paper deals with representation of physical environement in multi-agent simulators. We introduce the notion of model of space, defined as the computational structure which translates the spatial relations that exists between the simulated movings entities and which controls the dynamics of these relations.The discussion analyses the wide diversity of solutions actually adopted in simulators and classifies these various solutions in two main types of models. Then the paper comments on how these models should be integrated in the computer architecture of multi-agent simulators, as an independtly programmed toolbox.

We present a simulator designed to help scientists who are not experts in computing, such as biologists, to build up, run, and update their models in the field of fish population dynamics. This simulator, written in Smalltalk, is built upon multi-agents concepts and is well suited to model systems where individual or spatial aspects are involved. It provides tools to set the state and the behavior of the agents and to set their environment. The pre-defined components should preferably be used, but help for customizing components or programming new ones i s supplied, in order to build more complex agents. The simulator manages qualitative and quantitative variables with simple or composed units, dependencies on parameters, synchronous and asynchronous modes within and between agents, and the importation of ASCII files for the scenario scripts or for the initial stocking of the agents. Four examples of applications are presented.

In this paper a knowledge-based decision support system is described that determines the abiotic (chemical and physical) characteristics of a site on the basis of in-homogeneous samples of plant species. Techniques from the area of non-monotonic reasoning are applied to model multi-interpretable input information.

This paper describes a simulation environment, called Cormas, that relies on multi-agent systems and has been achieved in Smalffalk, using VisualWorks software. Such a simulation tool may prove useful to better understand the complex interactions between natural and social dynamics when studying renewable resource management. The general principles of the Cormas platform are first presented, then the implementation is described. Two models built with Cormas allow to illustrate the use and the genericity of this tool.

With the growing amount of distributed and heterogeneous information, conventional Environment Information Systems (EIS) have come to their limits. In order to overcome these problems, this work elaborates the concept of the Logical Client, an agent-enhanced architecture that is used by participants of the EIS. It consists of independent agents and uses artificial intelligence to cope with complex patterns of communication and actions. The Service Agent Layer as communication framework is based on CORBA, KQML and KIF. The standardised interface agents use rule based systems to find adequate reactions to incoming messages. The Strategy Service as an example of an agent is a utility-based agent for dynamic process modelling. It plans workflow processes depending on the needs of the user and the availability of services and resources. This architecture of the Logical Client has the potential to overcome complexity in information systems.

The landscape study is an interpretation task, usually performed in environmental planning, for which a wide range of expert knowledge sources is required. In order to help the experts in this field, a rule-based system that operates on fuzzy domains, has been used. However, some consistency problems together with an insufficient capability of adaptation to users' requirements have been detected. To improve this system, a new knowledge-based system has been designed and implemented. This approach exploits the fact that gradual knowledge rules are available, as these were obtained in the knowledge acquisition phase. In this work, we describe the main characteristics of such a system as well as its contribution to improve the capacity of adaptation to a great variety of user' inputs.

The problem of daily parking of subway vehicles is a special type of timetabling problems. It consists of assigning to the railroad lines of the depot of the subway system of Tunis the fleet of vehicles, while respecting a set of capacity and precedence constraints. Each vehicle is assigned daily one or more operating indices characterized by their departure and arrival times, and their itineraries. The objective is to minimize the number of scheduled conflicts. Five two-phase methods based on an exact model, on Genetic Algorithms, Tabu Search, Simulated Annealing and Expert Systems are proposed. The results are compared showing the superiority of the Tabu Search and Genetic Algorithm based heuristics and their capability of assigning all the vehicles in most cases.

In this paper the possibility of using the Artificial Neural Network technique for the individual Extensive Air Showers data evaluation is discussed. It is shown that the recently developed new computational methods can be used in studies of Extensive Air Showers registered by very large and complex detector systems. The Artificial Neural Network can be used to find a particular Extensive Air Shower parameter like e.g. total muou number. The example using the KASCADE array experiment geometry is given.

In consensus research, it is necessary to find criteria to assign confidence factors to the knowledge-based systems involved in a consensus algorithm. Those factors must reflect the confidence that we can have on each system's assertions. A whole class of such criteria are static ones (we call them quality criteria), that is, criteria based on the structure of the systems more than on any performance measure.In the present work, we propose, justify and formalize three static quality evaluation criteria for fuzzy systems: Completeness, Redundancy and Consistence. They are based on similar ones existing in classical logic, but they are generalized to the fuzzy domain. This is mainly done by making use of the subsethood theorem of Kosko's Set-as-Points framework, a very convenient way to assign geometric meaning to fuzzy sets.

Many types of earth retaining wall exist. The selection of the most appropriate type for a particular situation is not straightforward, and common practice is to select a type which the engineer has used previously elsewhere. This practice does not often lead to the most appropriate choice. The paper describes the development, use and evaluation of a prototype knowledge-based system to aid the selection of the most appropriate earth retaining wall. The working prototype system described in the paper is being extended. This is aided greatly by feedback from users and to encourage evaluative feedback, the software is made freely available.

The modernisation of the control systems in high-risk industrial environments presents a number of difficulties with regard to the reliability and security of the new implementation and especially with regard to the validation of the design and substitution process. In the particular case of a nuclear power plant, the specification of the original design is often not available, adding a further difficulty to the redesign process. In the context of system modernisation, the use of new implementation technologies enables learning possibilities to be incorporated in the redesigned systems. In this article we describe how to tackle the problem of the substitution of such control systems using a modular parametric architecture and we discuss the advantages of such an approach. We present, in generic form, how to approach the substitution of analogue control systems using this modular parametric architecture. We then apply this method to the particular case of the substitution of the system controlling the feed water circuit in a nuclear power plant and we give the experimental results for our implementation.

We modeled the esophageal myenteric nervous plexus as three morphological types of ganglia, arranged into neural chains and interconnected to form a network. Cholinergic synapses provide excitatory connections among ganglia and adrenergic synapses provide inhibitory activity within a ganglion. Mathematical formulation is based on HodgkinHuxley formalism for the dynamics of the nerve-pulse propagation along the axons and on Michaelis-Menten kinetics for the processes of electrochemical coupling at synapses. We studied the response of the neural network to external stimulation under “normal physiological” conditions and after simulated treatment with chloride salts of divalent ions, botulinum toxin, increased extracellular concentration of calcium ions and agents which alter the permeability of sodium and potassium channels. The special properties of the network tend to preserve nerve-signal conduction despite such noxious interventions.

When designing a technical system, simulation is an important concept to study the behavior of the planned system. Often a cycle of parameter variation and simulation is necessary to analyze the system behavior in detail or to improve the design of the system. Thus, the efficiency by which simulation can be carried out plays a role with respect to time, quality, and cost of the design process.The paper in hand shows in which way the simulation of technical systems can be speeded up. Starting point is the observation that for the different mathematical problems, which must be solved when simulating a system, several numerical methods are at hand. For instance, a system of ordinary differential equations can be solved by means of an explicit or an implicit Runge Kutta procedure.Since the different numerical methods are designed with respect to different qualities of a mathematical problem, there exists among the set of competing methods usually one suited best to do the required job. l. e., the qualities of a concrete mathematical problem can be used to select the best method. A central contribution of this paper is to show how this selection process can be operationalized.

In this paper a new fuzzy adaptive cancellation control scheme is presented. The basic part of the fuzzy adaptive cancellation controller is the inverse fuzzy model, which is given in the form of a fuzzy relational matrix. The fuzzy adaptive controller has been evaluated by implementation on heat-exchanger plant, which exhibits nonlinear and time-varying behaviour. To realize on-line identification, a recursive fuzzy identification algorithm based on the relational matrix has been developed. This identification algorithm offers very fast convergence of estimated parameters and the advantage of implicitly estimated operating point. It is shown that for mutable processes the adaptive fuzzy cancellation controller is superior to the classical model-reference adaptive control.