Skip to main content
main-content

Über dieses Buch

Anticipatory Learning Classifier Systems describes the state of the art of anticipatory learning classifier systems-adaptive rule learning systems that autonomously build anticipatory environmental models. An anticipatory model specifies all possible action-effects in an environment with respect to given situations. It can be used to simulate anticipatory adaptive behavior.

Anticipatory Learning Classifier Systems highlights how anticipations influence cognitive systems and illustrates the use of anticipations for (1) faster reactivity, (2) adaptive behavior beyond reinforcement learning, (3) attentional mechanisms, (4) simulation of other agents and (5) the implementation of a motivational module. The book focuses on a particular evolutionary model learning mechanism, a combination of a directed specializing mechanism and a genetic generalizing mechanism. Experiments show that anticipatory adaptive behavior can be simulated by exploiting the evolving anticipatory model for even faster model learning, planning applications, and adaptive behavior beyond reinforcement learning.

Anticipatory Learning Classifier Systems gives a detailed algorithmic description as well as a program documentation of a C++ implementation of the system.

Inhaltsverzeichnis

Frontmatter

Chapter 1. Background

Abstract
In order not to precipitately rush into the theoretical details and functioning of ACS2, we first provide a rather general background of anticipations, genetic algorithms, and learning classifier systems. ACS2 is certainly not a revolutionary new approach but rather a consequence and long awaited necessity in the research areas of anticipations, adaptive behavior, reinforcement learning, and learning classifier systems. The background section provides insights about what led to the consequent development of ACS2 and introduces all ideas relevant to the system.
Martin V. Butz

Chapter 2. ACS2

Abstract
The background chapter illustrated an important cognitive aspect of the current knowledge about learning. The insight of the presence and importance of anticipations in animals and man leads to the conclusion that it is mandatory to represent and learn anticipations also in animats. Furthermore, once this task is accomplished, it is necessary to investigate how the usefulness of anticipations can be exploited by the animat. The development of such a system and the exploitation of the anticipations in the resulting system is the enterprise of the remainder of this book.
Martin V. Butz

Chapter 3. Experiments with ACS2

Abstract
Although the genetic generalization approach in the last chapter reveals many similarities with XCS, it is certainly not straight forward to confirm that the GA is any good. The GA could be too strong, which would result in a disruption of the model evolution and an over-generalized population, or too weak, which would result in a larger, still over-specialized population. It is examined if the GA is able to overcome the observed over-specialization. Does the GA result in an evolution of accurate, maximally general classifiers in the population and further, does the GA cause a convergence to the aspired classifiers?
Martin V. Butz

Chapter 4. Limits

Abstract
The last chapter was intended to show the variety of problems in which ACS2 is able to learn. Model learning results have been provided in four different tasks showing that ACS2 is indeed capable of forming a generalized, complete model of an environment. Genetic generalization showed to be able to remedy the possibly occurring over-specializations in the ALP. The result was a decrease in population size and model size and a consequent decrease in computation time and the evolution of an even better environmental model. Moreover, the reinforcement learning capabilities of ACS2 were shown in the multiplexer and the maze environment.
Martin V. Butz

Chapter 5. Model Exploitation

Abstract
Despite the distinct challenges in the last chapter, the experimental studies confirmed that ACS2 is normally able to evolve a compact, complete, and accurate model of an environment. While the experiments in chapter 3 mainly showed the performance of ACS2 in terms of its model and reinforcement learning capabilities, this chapter exhibits experiments in which performance is increased by exploiting the environmental model. Particularly, promising results of two model-exploitation types are provided. (1) A complete and accu­rate model is evolved faster and more reliable. (2) A further adaptivity beyond the usual reinforcement learning capabilities is realized.
Martin V. Butz

Chapter 6. Related Systems

Abstract
The last chapters introduced and evaluated the ACS2 learning mechanism. The intensive investigation showed that ACS2 is able to tackle a large variety of problems ranging from hard classification problems to big environmental tasks. The genetic generalization mechanism confirmed to be capable to further generalize the evolving environmental model working together with the directed specialization caused by an anticipatory learning process (ALP). Several distinct challenges revealed the actual shortcomings in the system. Model exploitation processes were able to overcome several of the challenges as well as were able to further increase model learning performance and adaptivity. Model exploitation also showed the distinct adaptive behavior capabilites of the system. All in all, the last chapters provided a unified few of the functioning, capabilities, and limits of the current ACS2 system.
Martin V. Butz

Chapter 7. Summary, Conclusions, and Future Work

Abstract
This work introduced an anticipatory learning classifier system, namely ACS2. The system autonomously evolves an internal environmental model represented by a population of condition-action-effect classifiers. That is, without any supervision, ACS2 is acting upon an environment and is forming a model of the encountered environment meanwhile. Although the interaction is unsupervised, the resulting state after the execution of an action can be viewed as a supervision, so that the model learning mechanism might be termed implicitly supervised. After a summary of the presented work, this chapter concludes the book with a discussion of possible future work with ACS2.
Martin V. Butz

Backmatter

Weitere Informationen