Rule-based evolutionary online learning systems, often referred to as Michigan-style learning classifier systems (LCSs), were proposed nearly thirty years ago (Holland, 1976; Holland, 1977) originally calling them cognitive systems. LCSs combine the strength of reinforcement learning with the generalization capabilities of genetic algorithms promising a flexible, online generalizing, solely reinforcement dependent learning system. However, despite several initial successful applications of LCSs and their interesting relations with animal learning and cognition, understanding of the systems remained somewhat obscured. Questions concerning learning complexity or convergence remained unanswered. Performance in different problem types, problem structures, concept spaces, and hypothesis spaces stayed nearly unpredictable. This thesis has the following three major objectives: (1) to establish a facetwise theory approach for LCSs that promotes system analysis, understanding, and design; (2) to analyze, evaluate, and enhance the XCS classifier system (Wilson, 1995) by the means of the facetwise approach establishing a fundamental XCS learning theory; (3) to identify both the major advantages of an LCS-based learning approach as well as the most promising potential application areas. Achieving these three objectives leads to a rigorous understanding

Abstract. A new approach to building robots with anticipatory behavior is presented. This approach is based on analogy with a single episode from the past experience of the robot. The AMBR model of analogy-making is used as a basis, but it is extended with new agent-types and new mechanisms that allow anticipation related to analogical transfer. The role of selective attention on retrieval of memory episodes is tested in a series of simulations and demonstrates the context sensitivity of the AMBR model. The results of the simulations clearly demonstrated that endowing robots with analogy-based anticipatory behavior is promising and deserves further investigation. 1

Abstract. The goal of this paper is to analyse the central role of the epistemic activity in anticipatory mental life. A precise typology of different kinds of epistemic actions will be presented. On theses basis, a precise characterization of expectations about the success in the achievement of an intended result will be provided. Moreover, two specific kinds of epistemic actions (Epistemic Control and Epistemic Monitoring) will be defined and their functions in the goal-processing will be identified. 1

The cognitive abilities of the anticipatory classifier system (ACS) have already been successfully shown in earlier work (Stolzmann et al 2000). This report takes inspiration from some philosophical ideas for the purpose of dreaming in animals and humans during REM sleep. This is supported by recent neurological studies that show that rats revisit recent situations in a way that suggests dreaming (Wilson et al 2001). A simple extension is made to the ACS that uses the incomplete information contained in the classifier list as a basis for an abstract world model in which to interact or ’dream’. The abstract thread or dream direction is an emergent property of the selection process, this can be used to recycle around well known states and reduce real world interaction. The system is applied to two simple problems, the random walk and T−maze experiment and demonstrate that they require considerably less interactions with the real world to develop confident world models. Further models and extensions are proposed to advance the system, such as environmental directed generalisation and speculative rule creation.

Design and anticipation: towards an organisational view of design systems ISSN 1467-1298

Learning Classifier Systems (LCSs), such as XCS and other accuracy-based classifier systems, evolve distributed problem solutions represented by a population of rules. During evolution, features are specialized, propagated, and recombined to provide increasingly accurate subsolutions. Recently, it was shown that, like in conventional genetic algorithms (GAs), some problems require the processing of subsets of features as opposed to individual features to find the problem solution efficiently. In such problems, standard variation operators of genetic and evolutionary algorithms used in LCSs suffer from potential disruption of groups of interacting features, resulting in poor performance. This paper introduces two competent crossover operators to XCS by incorporating techniques derived from competent GAs: the extended compact GA (ECGA) and the Bayesian optimization algorithm (BOA). Instead of simple crossover operators such as uniform crossover or one-point crossover, here a probabilistic model of the global population is built and sampled to generate offspring classifiers locally. The distinction between the global and local problem structure is an additional challenge since the local problem structure may differ from the global structure. Two offspring generation methods are introduced and evaluated. The results