AI 2012: Advances in Artificial Intelligence

An important aim in bilateral negotiations is to achieve a win-win solution for both parties; therefore, a critical aspect of a negotiating agent’s success is its ability to take the opponent’s preferences into account. Every year, new negotiation agents are introduced with better learning techniques to model the opponent. Our main goal in this work is to evaluate and compare the performance of a selection of state-of-the-art online opponent modeling techniques in negotiation, and to determine under which circumstances they are beneficial in a real-time, online negotiation setting. Towards this end, we provide an overview of the factors influencing the quality of a model and we analyze how the performance of opponent models depends on the negotiation setting. This results in better insight into the performance of opponent models, and allows us to pinpoint well-performing opponent modeling techniques that did not receive much previous attention in literature.

We use optimism to introduce generic asymptotically optimal reinforcement learning agents. They achieve, with an arbitrary finite or compact class of environments, asymptotically optimal behavior. Furthermore, in the finite deterministic case we provide finite error bounds.

The primitive unicellular organism slime mould

Physarum Polycephalum

has attracted much attention from researchers of both biology and computer science fields. Biological experiments have revealed that its foraging mechanism can be used to solve shortest path problems, while its foraging process can construct efficient networks among food sources. Oregonator Model and Cellular Automaton have been proposed to simulate the intelligence and morphology of

Physarum

. To better understand the network formation of

Physarum

, a multi-agent system (MAS) model of particles was introduced by Jones, which can simulate many interesting patterns of

Physarum

transport networks. The MAS model is improved in three aspects: the number of sensors of each individual agent is reduced to two, while the function of each sensor is extended to sample both chemical nutrient and trail. A memory module is added to the architecture of an agent, by which the evolution mechanism can be achieved to maintain the population of the system. With such improvements, the system is more flexible and adaptive, and the networks constructed using the MAS model are more approximate to the ones by

Physarum

in biological experiments. All these are verified by constructing stable networks including Steiner’s minimum tree, cycle-like and spanning trees.

Real-time environmental monitoring can provide vital situational awareness for effective management of natural resources. Effective operation of Shellfish farms depends on environmental conditions. In this paper we propose a supervised learning approach to predict the farm closures. This is a binary classification problem where farm closure is a function of environmental variables. A problem with this classification approach is that farm closure events occur with small frequency leading to class imbalance problem. Straightforward learning techniques tend to favour the majority class; in this case continually predicting no event. We present a new ensemble class balancing algorithm based on random undersampling to resolve this problem. Experimental results show that the class balancing ensemble performs better than individual and other state of art ensemble classifiers. We have also obtained an understanding of the importance of relevant environmental variables for shellfish farm closure. We have utilized feature ranking algorithms in this regard.

Multilayer feedforward neural networks trained via supervised learning have proven to be successful in pattern recognition. This paper presents the technique of using single hidden layer feedforward neural network as an automatic classifier in music classification. Han Chinese folk songs from five distinct geographical regions in China are studied and encoded using a novel musical feature density map (MFDMap) for machine classification. The extreme learning machine (ELM) and its two variants are employed as the classifiers to categorize the folk songs. Our simulations show that by using a low-pass finite impulse response extreme learning machine (FIR-ELM), we can achieve 80.65% classification accuracy.

People-to-people recommendation aims at suggesting suitable matches to people in a way that increases the likelihood of a positive interaction. This problem is more difficult than conventional item-to-people recommendation since the preferences of both parties need to be taken into account. Previously we proposed a profile-based recommendation method that first uses compatible subgroup rules to select a single best attribute value for each corresponding value of the user, then combines these attribute value pairs into a rule that determines the recommendations. Though this method produces a significant improvement in the probability of an interaction being successful, it has two significant limitations: (i) by considering only single matching attribute values the method ignores cases where different attribute values are closely related, missing potential candidates, and (ii) when ranking candidates for recommendation the method does not consider individual behaviour. This paper addresses these two issues, showing how multiple attributes can be used with compatible subgroup rules and individual reply rates used for ranking candidates. Our experimental results show that the new approach significantly improves the probability of an interaction being successful compared to our previous approach.

Psoriasis is a common skin disease with no known cure. It is both subjective and time consuming to evaluate the severity of psoriasis lesions using manual methods. More objective automated methods are in great demand in both psoriasis research and in clinical practice. This paper presents an algorithm for scoring the severity of psoriasis lesions from 2D digital skin images. The algorithm uses the redness of the inflamed skin, or erythema, and the relative area and roughness of the flaky scaled skin, or scaling, in lesions to score lesion severity. The algorithm is validated by comparing the severity scores given by the algorithm against those given by dermatologists and against other automated severity scoring techniques.

This paper proposes a technique for the creation of a neural ensemble that introduces diversity through incorporating ten-fold cross validation together with varying the number of neurons in the hidden layer during network training. This technique is utilized to improve the classification accuracy of masses in digital mammograms. The proposed technique has been tested on a widely available benchmark database.

As lifelong learning becomes increasingly important in our society, mechanisms allowing students to evaluate their progress must be provided. A commonly used and widely accepted feedback mechanism is the multiple-choice test. Manual creation of multiple choice questions is often a time consuming process involving many iterations of trail and error. Using text processing and natural language processing techniques, automated multiple choice question generation, in recent years, is getting much closer to reality than ever. However, one of the most difficult tasks in both manual creation and automated generation of this kind of tests is the creation of distractors, because unsuitable distractors allow students to easily guess the correct answer, which counteracts the goal of these questions. In this paper, we investigated the desired properties of distractors and identified relevant text processing algorithms, specifically, latent semantic analysis and stylometry, for distractor selection. The refined distrators are compared with baseline distrators generated by our existing Automated Question Creator (AQC). Our preliminary evaluation shows that this novel combined approach produces distractors with a higher quality than those of the baseline AQC system.

In this paper, we present a new genetic algorithm for protein structure prediction problem using face-centred cubic lattice and hydrophobic-polar energy model. Our algorithm uses

i

) an exhaustive generation approach to diversify the search;

ii

) a novel hydrophobic core-directed macro move to intensify the search; and

iii

) a random-walk strategy to recover from stagnation. On a set of standard benchmark proteins, our algorithm significantly outperforms the state-of-the-art algorithms for the same models.

To combat online auction fraud, researchers have developed fraud detection and prevention methods. However, it is difficult to effectively evaluate these methods using commercial or synthetic auction data. For commercial data, it is not possible to accurately identify cases of fraud. For synthetic auction data, the conclusions drawn may not extend to the real world. The availability of realistic synthetic auction data, which models real auction data, will be invaluable for effective evaluation of fraud detection algorithms. We present an agent-based simulator that is capable of generating realistic English auction data. The agents and model are based on data collected from the TradeMe online auction site. We evaluate the generated data in two ways to show that it is similar to the TradeMe data. Evaluation of individual features show that correlation is greater than 0.9 for 8 of the 10 features, and evaluation using multiple features gives a median accuracy of 0.87.

This paper proposes a global motion estimation method to remove unintentional camera motions which degrade the visual quality of image sequences. The proposed approach is based on combination of 2D Radon transform, 1D Fourier transform and 1D Scale transform which can accurately estimate scale, rotational and translational distortions of camera motion and is robust to internal moving objects. Our experimental results with real and synthesized videos indicate the effectiveness of our proposed method.

This paper investigates automatic construction of invariant features using Genetic Programming (GP) for edge detection. Generally, basic features for edge detection, such as gradients, are further manipulated to improve detection performance. In order to improve detection performance, new features are constructed from different local features. In this study, GP is proposed to automatically construct invariant features based on basic invariant features from gradients, image quality (means and standard deviations), and histograms of images. The experimental results show that the invariant features constructed by GP combine advantages from the basic features, reduce drawbacks from basic features alone, and also improve the detection performance.

In order to allow more flexible and general learning, it is an advantage for artificial systems to be able to discover re-usable features that capture structure in the environment, known as Deep Learning. Techniques have been shown based on convolutional neural networks and stacked Restricted Boltzmann Machines, which are related to some degree with neural processes. An alternative approach using abstract representations, the ARCS Learning Classifier System, has been shown to build feature hierarchies based on reinforcement, providing a different perspective, however with limited classification performance compared to Artificial Neural Network systems. An Abstract Deep Network is presented that is based on ARCS for building the feature network, and introduces gradient descent to allow improved results on an image classification task. A number of implementations are examined, comparing the use of back-propagation at various depths of the system. The ADN system is able to produce classification error of 1.18% on the MNIST dataset, comparable with the most established general learning systems on this task. The system shows strong reliability in constructing features, and the abstract representation provides a good platform for studying further effects such as as top-down influences.

Significant progress has been made in automatic facial expression analysis using facial images and videos. The recognition reliability of most current approaches is still poor in naturalistic expressions compared to acted ones. Most of these methods use a static image of each expression that captures the characteristic image at the apex. However, according to psychologists, analyzing a sequence of images in a dynamic manner produces more accurate and robust recognition of facial affect expressions. In this paper, a new dynamic model is proposed for detecting naturalistic affect expressions. The Local Binary Pattern in Three Orthogonal Planes (LBP-TOP) is considered for modeling appearance and motion of facial features. The International Affective Picture System (IAPS) collection was used as stimulus for triggering naturalistic affective states. The dynamic approach produced an improvement of 16% for valence classification and 22% for arousal classification over previous studies.

Constrained optimization is an active area of research where attempts are being regularly made to improve the efficiency of the underlying optimization algorithms. While population based stochastic algorithms such as evolutionary algorithms, differential evolution (DE), particle swarm optimization etc. have been the popular choice as the underlying optimization scheme, adaptive strategies are usually employed to deal with constraints. Most of such approaches adopt a complete evaluation policy, i.e., all constraints and objectives corresponding to a solution is always evaluated for every solution under consideration. However, in a typical constrained optimization problem, one or more constraints are often difficult to satisfy and it might be beneficial to evaluate the constraints in a sequence. Evaluation of subsequent constraints and objective function can be skipped whenever a constraint is violated. In this paper, a self adaptive differential evolution algorithm is introduced which maintains multiple subpopulations, each of which is assigned a prescribed constraint sequence based on a ring topology. Solutions are ranked in each subpopulation and a migration scheme is employed to transfer feasible solutions to a subpopulation of feasible individuals. The performance of the proposed scheme is compared with a single sequence approach and other state of the art DE forms using the standard

g-series

test functions having inequality constraints. The results clearly indicate the potential savings in the computational cost. Apart from savings in computational cost, the paper also makes an important contribution as it provides useful physical insights on the search trajectories and their effect in various forms of constrained optimization problems.

Local Search approaches to constraint satisfaction are based on the ability to compute

violation

scores. These estimate ’how far’ a given assignment is from satisfying the constraints, and aim at guiding the search towards assignments whose violation will, ultimately, be null. We study the computation of violation functions for Boolean predicates encoded in

Decomposable Negation Normal Form

, an important class of logical circuits that encode interesting constraints. We show that for these circuits an elegant and efficient violation function first proposed by Z. Stachniak is ”ideal” in a certain sense. We discuss more broadly the notion of ”ideal” violation and discuss the implications of this result.

We study propagation of a global constraint that ensures that each row of a matrix of decision variables satisfies a

Regular

constraint, and each column satisfies a

Gcc

constraint. On the negative side, we prove that propagation is

NP

-hard even under some strong restrictions (e.g. just 2 values, just 4 states in the automaton, just 5 columns to the matrix, or restricting to limited classes of automata). We also prove that propagation is

W[2]

-hard when the problem is parameterized by the number of rows in the matrix. On the positive side, we identify several cases where propagation is fixed parameter tractable.

Stochastic perturbation on variable flipping is the key idea of local search for SAT. Observing that variables are flipped several times in an attempt to escape from a local minimum, this paper presents a duplication learning mechanism in stagnation stages to minimise duplicative variable flipping. The heuristic incorporates the learned knowledge into a variable weighting scheme to effectively prevent the search from selecting duplicative variables. Additionally, probability-based and time window smoothing techniques are adopted to eliminate the effects of redundant information. The integration of the heuristic and gNovelty

 + 

was compared with the original solvers and other state-of-the-art local search solvers. The experimental results showed that the new solver outperformed other solvers on the full set of SAT 2011 competition instances and three sets of real-world verification problems.

Heuristic search algorithms (MOA*, NAMOA* etc) for multiobjective optimization problems, when run with admissible heuristics, typically spend a lot of time and require huge amount of memory for generating the Pareto optimal solution frontier due to the fact that these algorithms expand all nodes/paths whose cost is not dominated by any other optimal solution. In this paper, we present an anytime heuristic search framework for biobjective optimization problems named “Anytime Biobjective Search (ABS)” which quickly finds a set of nondominated solutions and keeps on improving the solution frontier with time. The proposed framework uses the upper and lower limit estimates on one of the objectives to split the search space into a given number of segments and independently runs a particular search algorithm (branch-and-bound, beam search etc.) within each of the segments. In this paper, we present how existing search strategies, branch-and-bound, beam, and beam-stack, can be used within the proposed framework. Experimental results reveal that our proposed framework achieves good anytime performance.

Constraint satisfaction problem (CSP) is a subset of optimization problem where at least one solution is sought that satisfies all the given constraints. Presently, evolutionary algorithms (EAs) have become standard optimization techniques for solving unconstrained optimization problems where the problem is formalized for discrete or continuous domains. However, traditional EAs are considered ‘blind’ to constraint as they do not extract and exploit information from the constraints. A variation of EA – intelligent constraint handling for EA (ICHEA) proposed earlier models constraints to guide the evolutionary search to get improved and efficient solutions for continuous CSPs. As many real world CSPs have constraints defined in the form of discrete functions, this paper serves as an extension to ICHEA that reports its applicability for solving discrete CSPs. The experiment has been carried on a classic discrete CSP – the N-Queens problem. The experimental results show that extracting information from constraints and exploiting it in the evolutionary search makes the search more efficient. This provision is a problem independent formulation in ICHEA.

Anytime heuristic search algorithms are widely applied where best-first search algorithms such as A* require large or often unacceptable amounts of time and memory. Anytime algorithms produce a solution quickly and iteratively improve the solution quality. In this paper, we propose novel anytime heuristic search algorithms with a common underlying strategy called Column Search. The proposed algorithms are complete and guarantee to produce an optimal solution. Experimental results on sliding-tile puzzle problem, traveling salesman problem, and robotic arm trajectory planning problem show the efficacy of proposed methods compared to state-of-the-art anytime heuristic search algorithms.

Classification of mass spectrometry (MS) data is an essential step for biomarker detection which can help in diagnosis and prognosis of diseases. However, due to the high dimensionality and the small sample size, classification of MS data is very challenging. The process of biomarker detection can be referred to as feature selection and classification in terms of machine learning. Genetic programming (GP) has been widely used for classification and feature selection, but it has not been effectively applied to biomarker detection in the MS data. In this study we develop a GP based approach to feature selection, feature extraction and classification of mass spectrometry data for biomarker detection. In this approach, we firstly use GP to reduce the “redundant” features by selecting a small number of important features and constructing high-level features, then we use GP to classify the data based on selected features and constructed features. This approach is examined and compared with three well known machine learning methods namely decision trees, naive Bayes and support vector machines on two biomarker detection data sets. The results show that the proposed GP method can effectively select a small number of important features from thousands of original features for these problems, the constructed high-level features can further improve the classification performance, and the GP method outperforms the three existing methods, namely naive Bayes, SVMs and J48, on these problems.

Autonomous underwater vehicles (AUVs) are becoming an attractive option for increasingly complex and challenging underwater search and survey operations. To meet the emerging demands of AUV mission requirements, design and tradeoff complexities, there is an increasing interest in the use of multidisciplinary design optimization (MDO) strategies. While optimization techniques have been applied successfully to a wide range of applications spanning various fields of science and engineering, there is very limited literature on optimization of AUV designs. Presented in this paper is an evolutionary approach for the design optimization of AUVs using two stochastic, population based optimization algorithms. The proposed approach is capable of modelling and solving single and multi-objective constrained formulations of the AUV design problems based on different user and mission requirements. Two formulations of the AUV design problem are considered to identify designs with minimum drag and internal clash-free assembly. The flexibility of the proposed scheme and its ability to identify optimum preliminary designs are highlighted in this paper.

The importance of selecting relevant features for data modeling has been recognized already in machine learning. This paper discusses the application of an evolutionary-based feature selection method in order to generate input data for unsupervised learning in DARA (Dynamic Aggregation of Relational Attributes). The feature selection process which is based on the evolutionary algorithm is applied in order to improve the descriptive accuracy of the DARA (Dynamic Aggregation of Relational Attributes) algorithm. The DARA algorithm is designed to summarize data stored in the non-target tables by clustering them into groups, where multiple records stored in non-target tables correspond to a single record stored in a target table. This paper addresses the issue of optimizing the feature selection process to select relevant set of features for the DARA algorithm by using an evolutionary algorithm, which includes the evaluation of several scoring measures used as fitness functions to find the best set of relevant features. The results show the unsupervised learning in DARA can be improved by selecting a set of relevant features based on the specified fitness function which includes the measures of the dispersion and purity of the clusters produced.

This paper compares the implementations and performance of two computational methods, hierarchical clustering and a genetic algorithm, for inference of phylogenetic trees in the context of the artificial organism

Caminalcules

. Although these techniques have a superficial similarity, in that they both use agglomeration as their construction method, their origin and approaches are antithetical. For a small problem space of the original species proposed by Camin (1965) the genetic algorithm was able to produce a solution which had a lower Fitch cost and was closer to the theoretical evolution of

Caminalcules.

Unfortunately for larger problem sizes its time cost increased exponentially making the greedy directed search of the agglomerative clustering algorithm a more efficient approach.

Dimension reduction aims to remove unnecessary attributes from datasets to overcome the problem of “the curse of dimensionality”, which is an obstacle in classification. Based on the analysis of the limitations of the standard rough set theory, we propose a new dimension reduction approach based on binary particle swarm optimisation (BPSO) and probabilistic rough set theory. The new approach includes two new specific algorithms, which are

PSOPRS

using only the probabilistic rough set in the fitness function and

PSOPRSN

adding the number of attributes in the fitness function. Decision trees, naive Bayes and nearest neighbour algorithms are employed to evaluate the classification accuracy of the reduct achieved by the proposed algorithms on five datasets. Experimental results show that the two new algorithms outperform the algorithm using BPSO with standard rough set and two traditional dimension reduction algorithms. PSOPRSN obtains a smaller number of attributes than PSOPRS with the same or slightly worse classification performance. This work represents the first study on probabilistic rough set for for filter dimension reduction in classification problems.

Recent years have seen a resurgence of interest in evolving plastic neural networks for online learning. These approaches have an intrinsic appeal – since, to date, the only working example of general intelligence is the human brain, which has developed through evolution, and exhibits a great capacity to adapt to unfamiliar environments. In this paper we review prior work in this area – including problem domains and tasks, fitness functions, synaptic plasticity models and neural network encoding schemes. We conclude with a discussion of current findings and promising future directions, including incorporation of functional properties observed in biological neural networks which appear to play a role in learning processes, and addressing the “general” in general intelligence by the introduction of previously unseen tasks during the evolution process.

Coevolutionary algorithms are a special kind of evolutionary algorithm with advantages in solving certain specific kinds of problems. In particular, competitive coevolutionary algorithms can be used to study problems in which two sides compete against each other and must choose a suitable strategy. Often these problems are multimodal — there is more than one strong strategy for each side. In this paper, we introduce a scalable multimodal test problem for competitive coevolution, and use it to investigate the effectiveness of some common coevolutionary algorithm enhancement techniques.

Wilson extended XCS with interval based conditions to XCSR to handle real valued inputs. However, the possible actions must always be determined in advance. Yet domains such as robot control require numerical actions, so that neither XCS nor XCSR with their discrete actions can yield high performance. In the work presented here, genetic programming-based representation is used for the first time to compute continuous action in XCSR. This XCSR version has been examined on a simple one-dimensional but non-linear testbed problem – the “frog” problem – and compared with two continuous action based systems, GCS and XCSFCA. The proposed approach has consistently solved the frog problem and outperformed GCS and XCSFCA.

Shape representation and the associated morphing (repair) operators play a vital role in any shape optimization exercise. This paper presents a novel and an efficient methodology for morphing via smart repair of control points, wherein a repaired sequence of control points are generated. The repaired set of control points are then used to define the curve or the surface using a B-spline representation, while the control points themselves are optimized using a memetic algorithm. While the authors have already proposed the approach for 2D shape matching, this paper extends the approach to deal with 3D shape matching problems. Two 2D and one 3D examples have been presented to illustrate the performance of the proposed approach.

Parallel Linear Genetic Programming (PLGP) is an architecture that addresses instruction dependencies in Linear Genetic Programming (LGP). The Co-operative Coevolution (CC) methodology has previously been applied to PLGP but implementations have not been able to improve performance over vanilla PLGP. In this paper we present Hill Climbing Parallel Linear Genetic Programming (HC-PLGP) which uses a local search to discover effective combinations (blueprints) of partial solutions that are evolved in subpopulations. By introducing a new caching technique we can efficiently search over the subpopulations, and our improved fitness function combined with normalisation and blueprint elitism address some of the weaknesses of the previous approaches. Hill Climbing Parallel Linear Genetic Programming (HC-PLGP) is compared to three PLGP architectures over six datasets, and significantly outperforms them on two datasets, is comparable on three, and is slightly worse on one dataset.

In a real poker game, one player can take actions of different styles in different situations. In this paper, a novel method is proposed to quantify and model the opponent’s style in corresponding situation of a hand. Based on the proposed representation of Action Pair, the value of the style can be calculated and stored as “experience”. When making a decision, the specific style will be obtained from the “experience”. The style and the observable information will be used to estimate the value of the opponent’s hand. In experiments, the obtained “experience” validates the correctness of our assumption that a player does not show an invariable style in all situations. The experimental results show that the agent player using our method can predict the value of the opponent’s hand and earn more money in fixed hands comparing with the original agent.

Interactive training is well suited to computer games as it allows game designers to interact with otherwise autonomous learning algorithms. This paper investigates the outcome of a group of five commercial first person shooter game designers using a custom built interactive training tool to train first person shooter bots. The designers are asked to train a bot using the tool, and then comment on their experiences. The five trained bots are then pitted against each other in a deathmatch scenario. The results show that the training tool has potential to be used in a commercial environment.

In real life games, a player’s belief about the consequence of a strategy is often ambiguous due to out-of-control factors in the environment where the games are played. However, existing work cannot handle this situation. To address the issue, we introduce a new kind of games, called ambiguous games, and incorporate human cognitive factors of ambiguity aversion and minimising regret to propose a concept of solution to such a game. Moreover, we also study how ambiguity degrees of belief about payoffs impact the outcomes of a game, and find the condition under which a player should release more or less ambiguous information to his opponents.

This research introduces a new query expansion method that uses Wikipedia and its hyperlink structure to find related terms for reformulating a query. Queries are first understood better by splitting into query aspects. Further understanding is gained through measuring how well each aspect is represented in the original search results. Poorly represented aspects are found to be an excellent source of query improvement. Our main contribution is the way of using Wikipedia to identify aspects and underrepresented aspects, and to weight the expansion terms. Results have shown that our approach improves the original query and search results, and outperforms two existing query expansion methods.

In this paper, we propose an algorithm encouraging group sparsity under some convex constraint. It stems from some applications where the regression coefficients are subject to constraints, for example nonnegativity and the explanatory variables are not suitable to be orthogonalized within groups. It takes the form of the group LASSO based on linear regression model where a L1/L2 norm is imposed on group coefficients to achieve group sparsity. It differs from the original group LASSO in the following ways. First, the regression coefficients must obey some convex constraints. Second, there is no requirement for orthogonality of the variables within individual groups. For these reasons, the simple blockwise coordinate descent for all group coefficients is no longer applicable and a special treatment for the constraint is necessary. The algorithm we proposed in this paper is an alternating direction method, and both exact and inexact solutions are provided. The inexact version simplifies the computation while retaining practical convergence. As an approximation to group L0, it can be applied to data analysis where group fitting is essential and the coefficients are constrained. It may serve as a screening procedure to reduce the number of the groups when the number of total groups is prohibitively high.

When processing a noisy corpus such as clinical texts, the corpus usually contains a large number of misspelt words, abbreviations and acronyms while many ambiguous and irregular language usages can also be found in training data needed for supervised learning. These are two frequent kinds of noise that can affect the overall performance of machine learning process. The first noise is usually filtered by the proof reading process. This paper proposes an algorithm to deal with noisy training data problem, for a method we call reverse active learning to improve performance of supervised machine learning on clinical corpora. The effects of reverse active learning are shown to produce results on the i2b2 clinical corpus that are state-of-the-art of supervised learning method and offer a means of improving all processing strategies in clinical language processing.

Relevance feature and ontology are two core components to learn personalized ontologies for concept-based retrievals. However, how to associate user native information with common knowledge is an urgent issue. This paper proposes a sound solution by matching relevance feature mined from local instances with concepts existing in a global knowledge base. The matched concepts and their relations are used to learn personalized ontologies. The proposed method is evaluated elaborately by comparing it against three benchmark models. The evaluation demonstrates the matching is successful by achieving remarkable improvements in information filtering measurements.

Increased attention has been focused on question answering (QA) technology as next generation search since it improves the usability of information acquisition from web. However, not much research has been conducted on “non-factoid-QA”, especially on

Why Question Answering

(

Why-QA

). In this paper, we introduce a machine learning approach to automatically construct a classifier with function words as features to perform

Why Text Segments Classification

(

WTS

classification) by using SVM. It is a process of detecting text segments describing

“reasons-causes”

and is a subtask of

Why-QA

mainly related to an answer extraction part. We argue that function words are a strong discriminator for

WTS

classification. Furthermore, since function words appear in almost all text segments regardless of the domain of the topic, it also enables construction of a domain independent classifier. The experimental results showed significant improvement over state-of-the-art results in terms of accuracy of

WTS

classification as well as domain independent capability.

Meaning-preserving Skolemization is essential for development of a correct and efficient method of solving query-answering problems. It requires global existential quantifications of function variables, which in turn require an extended space of logical formulas. This paper proposes a bottom-up procedure for computing a set of models that sufficiently represents the set of all models of a given clause set in the extended formula space. Instantiations of function variables often result in generation of infinitely many models. To overcome the difficulty, a model-making pattern is introduced for representing a possibly infinite number of models, and such a pattern is split as late as possible. The proposed procedure provides a method for solving query-answering problems that include unrestricted use of universal and existential quantifications.

Unification in Description Logics (DLs) has been proposed as an inference service that can, for example, be used to detect redundancies in ontologies. For the DL

$\mathcal{EL}$

, which is used to define several large biomedical ontologies, unification is

NP

-complete. A goal-oriented NP unification algorithm for

$\mathcal{EL}$

that uses nondeterministic rules to transform a given unification problem into solved form has recently been presented. In this paper, we extend this goal-oriented algorithm in two directions: on the one hand, we add general concept inclusion axioms (GCIs), and on the other hand, we add role hierarchies (

$\mathcal{H}$

) and transitive roles (

R

 + 

). For the algorithm to be complete, however, the ontology consisting of the GCIs and role axioms needs to satisfy a certain cycle restriction.

One of the most successful approaches to the formalization of commonsense reasoning is the work by Lehmann and colleagues, known as the KLM approach, in which defeasible consequence relations with a preferential semantics are studied. In spite of its success, KLM is limited to propositional logic. In recent work we provided the semantic foundation for extending defeasible consequence relations to modal logics and description logics. In this paper we continue that line of investigation by going beyond the basic (propositional) KLM postulates, thereby making use of the additional expressivity provided by modal logic. In particular, we show that the additional constraints we impose on the preferential semantics ensure that the rule of necessitation holds for the corresponding consequence relations, as one would expect it to. We present a representation result for this tightened framework, and investigate appropriate notions of entailment in this context — normal entailment, and a rational version thereof.

This paper examines how Spohn’s

Ordinal Conditional Functions

, which are used to model belief dynamics, when appropriately adapted and interpreted, can fruitfully model the dynamics of trust. In this framework, trust is defined in terms of lack of trust (trust deficit) which is taken to be a primitive notion. It explores the notion of context sensitive trust, i.e., how in different domains of human interaction, one agent might trust another agent in varying degrees, and suggests the dynamics of such trust.

After an introduction to Heuristic-Driven Theory Projection (HDTP) as framework for computational analogy-making, and a compact primer on parametrized complexity theory, we provide a complexity analysis of the key mechanisms underlying HDTP, together with a short discussion of and reflection on the obtained results. Amongst others, we show that restricted higher-order anti-unification as currently used in HDTP is

W

[1]-hard (and thus

NP

-hard) already for quite simple cases. Also, we obtain W[2]-hardness, and

NP

-completeness, for the original mechanism used for reducing second-order to first-order anti-unifications in the basic version of the HDTP system.

Sensing actions, which allow an agent to increase its knowledge about the environment, are problematic for traditional planning languages. In this paper we propose a very general framework for representing both changes to the real world and to the knowledge of an agent, based on a first order linear time calculus. Our framework is more general than most existing approaches, because our semantics explicitly represents, for each point in time, not only the agent’s knowledge about that timepoint, but also about the past and the future. By applying a general approximation method for classical logic to this framework, we obtain an efficient and sound but incomplete reasoning method.

Human cognition is a mixture of the systematic and the non-systematic. One thing we can do systematically can be described as follows. If we know about multiplication, and the facts of basic multiplication, and we know conceptually what division is, then we can utilise the facts of multiplication that we know in order to solve division problems that correspond to those facts. For example, once children know that 4 ×7 = 28, and once they understand about division, they can work out that 28 / 4 = 7. Aizawa has defined standards for what counts as an explanation of systematicity. In this paper, in accordance with Aizawa’s framework, we apply concepts from category theory to this problem, and resolve it by identifying the unique natural transformation that underpins this example of systematicity, and others in the same class.

In this paper, we propose Rule Description Logic (RDL) for enhancing Description Logic (DL) with nonmonotonic recursive rules, like those in Answer Set Programming (ASP). We define the world view semantics for RDL and show that it is faithful with respect to both DL and ASP. More importantly, we show that the full language of RDL is decidable.

Outlier detection is an important process for text document collections, but as the collection grows, the detection process becomes a computationally expensive task. Random projection has shown to provide a good fast approximation of sparse data, such as document vectors, for outlier detection. The random samples of Fourier and cosine spectrum have shown to provide good approximations of sparse data when performing document clustering. In this article, we investigate the utility of using these random Fourier and cosine spectral projections for document outlier detection. We show that random samples of the Fourier spectrum for outlier detection provides better accuracy and requires less storage when compared with random projection. We also show that random samples of the cosine spectrum for outlier detection provides similar accuracy and computational time when compared with random projection, but requires much less storage.

The contextual bandit problem has been studied in the recommender system community, but without paying much attention to the contextual aspect of the recommendation. We introduce in this paper an algorithm that tackles this problem by modeling the Mobile Context-Aware Recommender Systems (MCRS) as a contextual bandit algorithm and it is based on dynamic exploration/exploitation. Within a deliberately designed offline simulation framework, we conduct extensive evaluations with real online event log data. The experimental results and detailed analysis demonstrate that our algorithm outperforms surveyed algorithms.

In some regression problems, it may be more reasonable to predict intervals rather than precise values. We are interested in finding intervals which simultaneously for all input instances

$x \in \mathcal{X}$

contain a

β

proportion of the response values. We name this problem simultaneous interval regression. This is similar to simultaneous tolerance intervals for regression with a high confidence level

γ

 ≈ 1 and several authors have already treated this problem for linear regression. Such intervals could be seen as a form of confidence envelop for the prediction variable given any value of predictor variables in their domain. Tolerance intervals and simultaneous tolerance intervals have not yet been treated for the

K

-nearest neighbor (KNN) regression method. The goal of this paper is to consider the simultaneous interval regression problem for KNN and this is done without the homoscedasticity assumption. In this scope, we propose a new interval regression method based on KNN which takes advantage of tolerance intervals in order to choose, for each instance, the value of the hyper-parameter

K

which will be a good trade-off between the precision and the uncertainty due to the limited sample size of the neighborhood around each instance. In the experiment part, our proposed interval construction method is compared with a more conventional interval approximation method on six benchmark regression data sets.

Although frequent pattern mining techniques have been extensively studied, the extension of their application onto data streams has been challenging. Due to data streams being continuous and unbounded, an efficient algorithm that avoids multiple scans of data is needed. In this paper we propose Kernel-Tree (KerTree), a single pass tree structured technique that mines frequent patterns in a data stream based on forecasting the support of current items in the future state. Unlike previous techniques that build a tree based on the support of items in the previous block, KerTree performs an estimation of item support in the next block and builds the tree based on the estimation. By building the tree on an estimated future state, KerTree effectively reduces the need to restructure for every block and thus results in a better performance and mines the complete set of frequent patterns from the stream while maintaining a compact structure.

In this paper we provide empirical data of the performance of the two most commonly used multiobjective reinforcement learning algorithms against a set of benchmarks. First, we describe a methodology that was used in this paper. Then, we carefully describe the details and properties of the proposed problems and how those properties influence the behavior of tested algorithms. We also introduce a testing framework that will significantly improve future empirical comparisons of multiobjective reinforcement learning algorithms. We hope this testing environment eventually becomes a central repository of test problems and algorithms The empirical results clearly identify features of the test problems which impact on the performance of each algorithm, demonstrating the utility of empirical testing of algorithms on problems with known characteristics.

Mining time series data and imbalanced data are two of ten challenging problems in data mining research. Imbalanced time series classification (ITSC) involves these two challenging problems, which take place in many real world applications. In the existing research, the structure-preserving over-sampling (SOP) method has been proposed for solving the ITSC problems. It is claimed by its authors to achieve better performance than other over-sampling and state-of-the-art methods in time series classification (TSC). However, it is unclear whether an under-sampling method with various learning algorithms is more effective than over-sampling methods, e.g., SPO for ITSC, because research has shown that under-sampling methods are more effective and efficient than over-sampling methods. We propose a comparative study between an under-sampling method with various learning algorithms and over-sampling methods, e.g. SPO. Statistical tests, the Friedman test and post-hoc test are applied to determine whether there is a statistically significant difference between methods. The experimental results demonstrate that the under-sampling technique with KNN is the most effective method and can achieve results that are superior to the existing complicated SPO method for ITSC.

Traditional classification methods assume that the training and the test data arise from the same underlying distribution. However in some adversarial settings, the test set can be deliberately constructed in order to increase the error rates of a classifier. A prominent example is email spam where words are transformed to avoid word-based features embedded in a spam filter. Recent research has modeled interactions between a data miner and an adversary as a sequential Stackelberg game, and solved its Nash equilibrium to build classifiers that are more robust to subsequent manipulations on training data sets. However in this paper we argue that the

iterative

algorithm used in the Stackelberg game, which solves an optimization problem at each step of play, is sufficient but not necessary for achieving Nash equilibria in classification problems. Instead, we propose a method that transforms singular vectors of a training data matrix to simulate manipulations by an adversary, and from that perspective a Nash equilibrium can be obtained by solving a novel optimization problem only

once

. We show that compared with the

iterative

algorithm used in recent literature, our

one-step

game significantly reduces computing time while still being able to produce good Nash equilibria results.

This paper presents a method for signals and features selection when classifying sleep apnea. This study uses a novel hierarchical parallel particle swarm optimization structure as proposed by the authors previously. In this structure, the swarms are separated into ‘masters’ and ‘slaves’ and access to global information is restricted according to their types. This method is used to classify sleep apneic events into apnea or hypopnea. In this study, ten different biosignals are used as the inputs for the system albeit with different features. The most important signals are subsequently determined based on their contribution to classification of the sleep apneic events. The classification method consists of three main parts which are: feature generation, signal selection, and data reduction based on PSO-SVM, and the final classifier. This study can be useful for selecting the best subset of input signals for classification of sleep apneic events, by attention to the trade of between more accuracy of higher number of input signals and more comfortable of less signals for the patient.

This paper examines the problem of modelling continuous, positive data by finite mixtures of inverse Gaussian distributions using the minimum message length (MML) principle. We derive a message length expression for the inverse Gaussian distribution, and prove that the parameter estimator obtained by minimising this message length is superior to the regular maximum likelihood estimator in terms of Kullback–Leibler divergence. Experiments on real data demonstrate the potential benefits of using inverse Gaussian mixture models for modelling continuous, positive data, particularly when the data is concentrated close to the origin or exhibits a strong positive skew.

This paper proposes a novel approach to solve the ordinal regression problem using Gaussian processes. The proposed approach, probabilistic least squares ordinal regression (PLSOR), obtains the probability distribution over ordinal labels using a particular likelihood function. It performs model selection (hyperparameter optimization) using the leave-one-out cross-validation (LOO-CV) technique. PLSOR has conceptual simplicity and ease of implementation of least squares approach. Unlike the existing Gaussian process ordinal regression (GPOR) approaches, PLSOR does not use any approximation techniques for inference. We compare the proposed approach with the state-of-the-art GPOR approaches on some synthetic and benchmark data sets. Experimental results show the competitiveness of the proposed approach.

Bagging ensemble selection (BES) is a relatively new ensemble learning strategy. The strategy can be seen as an ensemble of the ensemble selection from libraries of models (ES) strategy. Previous experimental results on binary classification problems have shown that using random trees as base classifiers, BES-OOB (the most successful variant of BES) is competitive with (and in many cases, superior to) other ensemble learning strategies, for instance, the original ES algorithm, stacking with linear regression, random forests or boosting. Motivated by the promising results in classification, this paper examines the predictive performance of the BES-OOB strategy for regression problems. Our results show that the BES-OOB strategy outperforms Stochastic Gradient Boosting and Bagging when using regression trees as the base learners. Our results also suggest that the advantage of using a diverse model library becomes clear when the model library size is relatively large. We also present encouraging results indicating that the non-negative least squares algorithm is a viable approach for pruning an ensemble of ensembles.

Applications of unsupervised learning techniques to action recognition have proved highly competitive in comparison to supervised and hand-crafted approaches, despite not being designed to handle image processing problems. Many of these techniques are either based on biological models of cognition or have responses that correlate to those observed in biological systems. In this study we apply (for the first time) an adaptation of the latest hierarchical temporal memory (HTM) cortical learning algorithms (CLAs) to the problem of action recognition. These HTM algorithms are both unsupervised and represent one of the most complete high-level syntheses available of the current neuroscientific understanding of the functioning of neocortex.

Specifically, we extend the latest HTM work on augmented spatial pooling, to produce a fixed frame temporal pooler (FFTP). This pooler is evaluated on the well-known KTH action recognition data set and in comparison with the best performing unsupervised learning algorithm for bag-of-features classification in the area: independent subspace analysis (ISA). Our results show FFTP comes within 2% of ISA’s performance and outperforms other comparable techniques on this data set. We take these results to be promising, given the preliminary nature of the research and that the FFTP algorithm is only a partial implementation of the proposed HTM architecture.

This paper introduces a new domain-independent keyphrase extraction system (DIKEA). Keyphrase extraction is a challenging problem that automatically extracts or assigns keyphrases to documents and it can benefit many research areas such as information retrieval, particularly indexing, clustering, and summarization. A landmark research KEA (Keyphrase Extraction Algorithm) formulated the problem as a supervised machine learning problem and successfully applied a Naïve Bayes model to it, which showed great promise but the performance is not satisfactory. Its state-of-the-art extension KEA++ has a significantly improved performance but relies on a domain specific vocabulary which is often not available or not complete. This paper introduces a novel domain-independent approach and has three main contributions: utilising the largest online knowledge source—Wikipedia—for keyphrase candidate selection; presenting new features for keyphrase evaluation, including a Wikipedia-based feature–link probability; and evaluating a number of different learning algorithms, including multilayer perceptrons, for keyphrase selection. Experiments show that our system clearly outperforms KEA and closely matches the performance of KEA++, without requiring any domain-specific knowledge such as KEA++’s vocabulary list.

Detecting events of interest in sensor data is crucial in many areas such as medical monitoring by body sensors. Current methods often require prior domain knowledge to be available. Moreover, it is difficult for them to find complex temporal patterns existing in multi-channel data. To overcome these drawbacks, we propose a Genetic Programming (GP) based event detection methodology which can directly take raw multi-channel data as input. By applying it to three event detection tasks with various event sizes and comparing with four typical classification methods, we can see that those detectors evolved by GP can handle raw data much better than other methods. With features manually defined based on domain knowledge, our method can also be comparable with others. The analysis of evolved detectors demonstrates that distinctive characteristics of the target events are captured by these GP detectors.

We study the problem of classification on uncertain objects whose locations are uncertain and described by probability density functions (pdf). Though there exist some classification algorithms proposed to handle uncertain objects, all existing algorithms are complex and time consuming. Thus, a novel supervised UK-means algorithm is proposed to classify uncertain objects more efficiently. Supervised UK-means assumes the classes are well separated. However, in real data, subsets of objects of the same class are usually interspersed among (disconnected by) other classes. Thus, we proposed a new algorithm Supervised UK-means with Multiple Subclasses (SUMS) which considers the objects in the same class can be further divided into several groups (subclasses) within the class and tries to learn the subclass representatives to classify objects more accurately.

Logistics and distribution planning are an important element of military operational planning. The vast level of detail available during planning increases the difficulty in forming a complete solution within time constraints. Moreover, the consideration of multiple planning scenarios is paramount to successfully investigate all available contingencies. This paper presents a hierarchical multi-agent approach to distribution planning. The use of a multi-agent system enables a distribution problem to be logically separated, with each part being delegated to an independent agent. The application of hierarchy allows agent communication to be regulated, reducing redundancy in communication, as encountered in a flat multi-agent structure. Through multiple trials, it was found that the application of hierarchy vastly improved the computational performance of the system, without compromise to solution quality.

Operations of modern organizations critically depend on Data Centers (DC). Due to ad hoc additions from diverse business units over time, the IT resources in a DC get unwieldy and complex. Transformations of DC - server consolidation, migration, application/data simplification, technology standardization - are important for cost, efficiency and reliability. Even when a specific transformation is identified (“consolidate these 100 existing servers into these 48 new servers”) it is difficult to generate a detailed optimal project plan for its execution. The project plan must identify all the tasks involved, identify an optimal team (size and expertise) and generate a detailed work schedule that meets the and respects the constraints and dependencies among the tasks. We present a methodology to generate such a plan automatically from given ”high-level” IT transformation specifications (“as-is” and “to-be” states). We adopt a heuristic forward chaining metric temporal planner engine (SAPA) to generate a project plan that attempts to optimize the overall time and team-size. The idea is to capture the domain-knowledge as reusable planning action. This automation reduces the efforts and errors in manual project planning. The method can be extended to projects in other domains.

We introduce a new class of planning problems in which there is a separate set of actions with higher priority than regular actions. We present new planning domains to show that problems of practical interest may easily fit in this framework. We argue that though this framework is quite succinctly encoded in classical planning itself, existing planners are disappointingly inept at solving them. To demonstrate this, we have built a wrapper tool for planners which uses ad-hoc techniques to give far better results. Therefore, we also propose our encoded domains as new challenges for general-purpose planners.

Capacitated Arc Routing Problem (CARP) is a well known combinatorial problem that requires the identification of the minimum total distance travelled by a set of vehicles to service a given set of roads subject to the vehicle’s capacity constraints. While a number of optimization algorithms have been proposed over the years to solve CARP problems, all of them require a large number of function evaluations prior to its convergence. Application of such algorithms are thus limited for practical applications as many of such applications require an acceptable solution within a limited time frame, e.g., dynamic versions of the problem. This paper is a pre-cursor to such applications, and the aim of this study is to develop an algorithm that can solve such problems with a limited computational budget of 50,000 function evaluations. The algorithm is embedded with a similarity based parent selection scheme inspired by the principles of multiple sequence alignment, hybrid crossovers, i.e., a combination of similarity preservation schemes, path scanning heuristics and random key crossovers. The performance of the algorithm is compared with a recent Memetic algorithm, i.e., Decomposition-Based Memetic Algorithm proposed in 2010 across three sets of commonly used benchmarks (

gdb

,

val

,

egl

). The results clearly indicate the superiority of performance across both small and large instances.

Partially ordered plans have several useful properties, such as exhibiting the structure of the plan more clearly which facilitates post-plan generation tasks like scheduling the plan, explaining it to a user, or breaking it into subplans for distributed execution. The standard interpretation of partial ordering implies that whenever two subplans are unordered, every interleaving of steps from the two forms a valid execution. This restricts deordering to cases where individual steps (i.e., actions) are independent. We propose a weaker notion of partial ordering that divides the plan into

blocks

, such that the steps in a block may not be interleaved with steps outside the block, but unordered blocks can be executed in any sequence. We present an algorithm to find such deorderable blocks, and show that it enables deordering plans in many cases where no deordering is possible under the standard interpretation.

Ride-sharing is considered as one of the promising solutions for reducing fuel consumption of fuel and reducing the congestion in urban cities, hence reducing the environmental pollution. With the advancement of mobile social networking technologies, it is necessary to reconsider the principles and desired characteristics of ride-sharing systems. Ride-sharing systems can be popular among people if we can provide more flexible and adaptive solution according to preferences of the participants and solve the social challenges. In this paper, we focus on encouraging people to use a ride-sharing system by satisfying their demands in terms of safety, privacy, convenience and also provide enough incentives for drivers and riders. We formalized the ride sharing problem as a multi source-destination path planning problem. An objective function is developed which models different conflicting objectives in a unified framework. We provide the flexibility to each driver that he can generate the sub-optimal paths according to his own requirements by suitably adjusting the weights. These sub-optimal paths are generated in an order of priority (optimality). The simulation results have shown that the system has the potential to compute multiple optimal paths.

The ability to accurately track a ball is a critical issue in humanoid robot soccer, made difficult by processor limitations and resultant inability to process all available data from a high-definition image. This paper proposes a computationally efficient method of determining position and size of balls in a RoboCup environment, and compares the performance to two common methods: one utilising Levenberg-Marquardt least squares circle fitting, and the other utilising a circular Hough transform. The proposed method is able to determine the position of a non-occluded tennis ball with less than 10% error at a distance of 5 meters, and a half-occluded ball with less than 20% error, overall outperforming both compared methods whilst executing 300 times faster than the circular Hough transform method. The proposed method is described fully in the context of a colour based vision system, with an explanation of how it may be implemented independent of system paradigm. An extension to allow tracking of multiple balls utilising unsupervised learning and internal cluster validation is described.

Recent research has shown the benefits of using

K

-means clustering in task allocation to robots. However, there is little evaluation of other clustering techniques. In this paper we compare

K

-means clustering to single-linkage clustering and consider the effects of straight line and true path distance metrics in cluster formation. Our empirical results show single-linkage clustering with a true path distance metric provides the best solutions to the multi-robot task allocation problem when used in sequential single-cluster auctions.

Without a model the application of reinforcement learning to control a dynamic system can be hampered by several shortcomings. The number of trials needed to learn a good policy can be costly and time consuming for robotic applications where data is gathered in real-time. In this paper we describe a variable resolution model-based reinforcement learning approach that distributes sample points in the state-space in proportion to the effect of actions. In this way the base learner economises on storage to approximate an effective model. Our approach is conducive to including background knowledge to speed up learning. We show how different types of background knowledge can used to speed up learning in this setting. In particular, we show good performance for a weak type of background knowledge by initially overgeneralising local experience.

In real life, sometimes Multi-Criteria Decision Making (MCDM) problems are dealt with inevitably under the disturbance of subjective factors such as intuition, feeling and emotion. However, the existing work cannot handle the issue well. As a result, it is difficult for them to predict and give some decision support for MCDM problem of this kind. Thus, to address the issue, this paper constructs a novel Analytic Hierarchy Process (AHP) approach based on Dempster-Shafer (D-S) theory. More specifically, our model can: (i) handle the subjective factor disturbances by distinguish complete (there not exists any hidden subjective factors impact the decision) and incomplete (the converse of complete) criteria; (ii) differentiate incomplete and complete relative ranking of the groups of decision alternatives over a criterion; and (iii) handle the ambiguous criteria evaluations of the groups of decision alternatives. Moreover, we give two methods to reduce the subjective factor disturbances. Finally, we illustrated our approach with a real-life problem.

Parameters in logistic regression models are commonly estimated by the method of maximum likelihood, while the model structure is selected with stepwise regression and a model selection criterion, such as AIC or BIC. There are two important disadvantages of this approach: (1) maximum likelihood estimates are biased and infinite when the data is linearly separable, and (2) the AIC and BIC model selection criteria are asymptotic in nature and tend to perform well only when the sample size is moderate to large. This paper introduces a novel criterion, based on the Minimum Message Length (MML) principle, for parameter estimation and model selection of logistic regression models. The new criterion is shown to outperform maximum likelihood in terms of parameter estimation, and outperform both AIC and BIC in terms of model selection using both real and artificial data.

In various domains, such as security and surveillance, a large amount of information from heterogeneous sources is continuously gathered to identify and prevent potential threats, but it is unknown in advance what the observed entity of interest should look like. The quality of the decisions made depends, of course, on the quality of the information they are based on. In this paper, we propose a novel method for assessing the quality of information taking into account uncertainty. Two properties –

soundness

and

completeness

– of the information are used to define the notion of information quality and their expected values are defined using a probabilistic model output. Simulation experiments with data from a maritime scenario demonstrates the usage of the proposed method and its potential for decision support in complex tasks such as surveillance.

While a great variety of algorithms have been developed and applied to learning static Bayesian networks, the learning of dynamic networks has been relatively neglected. The causal discovery program CaMML has been enhanced with a highly flexible set of methods for taking advantage of prior expert knowledge in the learning process. Here we describe how these representations of prior knowledge can be used instead to turn CaMML into a promising tool for learning dynamic Bayesian networks.