Evolution of cartesian genetic programs capable of learning
Gul Muhammad Khan, Julian Francis Miller · 2009
We propose a new form of Cartesian Genetic Programming (CGP) that develops into a computational network capable of learning. The developed network architecture is inspired by the brain. When the genetically encoded programs are run, a networks develops consisting of neurons, dendrites, axons, and synapses which can grow, change or die. We have tested this approach on the task of learning how to play checkers. The novelty of the research lies mainly in two aspects: Firstly, chromosomes are evolved that encode programs rather than the network directly and when these programs are executed they build networks which appear to be capable of learning and improving their performance over time solely through interaction with the environment. Secondly, we show that we can obtain learning programs much quicker through co-evolution in comparison to the evolution of agents against a minimax based checkers program. Also, co-evolved agents show significantly increased learning capabilities compared to those that were evolved to play against a minimax-based opponent.