Anticipatory learning classifier systems (Q4328477)

Anticipations are an important part of human behaviour. Anticipatory Learning Classifier Systems intend to apply this insight by simulation and utilization of anticipations in an artificial behavioural learning system. The book introduces such a system, namely ACS2. The system autonomously evolves an internal environmental model represented by a population of condition-action-effect classifiers. ACS2 is acting on an environment and is forming a model of the encountered environment. In seven chapters (Background, ACS2, Experiments with ACS2, Limits, model exploitation, related systems, concluding summary) it is shown how ACS2 evolves a complete, accurate, and compact environmental model, i.e. a model that anticipates the perceptual effects of an action in each situation accurately with a least number of maximally general classifiers. The model is built by two distinct processes: an anticipatory learning process and a genetic generalization. The two processes achieve the evolution of the intended environmental model. Despite some small over-generalization cases caused by the genetic generalization process, the extensive experimental study confirmed that the model evolves very reliably and genetic generalization results in great generalization capabilities. The final part discusses future work with ACS2. The appendix contains a detailed algorithmic description of ACS2 and a C++ program code documentation to investigate own experiments with anticipatory learning classifier systems.