Multi-behaviour robot control using genetic network programming with fuzzy reinforcement learning

This research explores a new hybrid evolutionary learning methodology for multi-behaviour robot control. The new approach is an extension of the Fuzzy Genetic Network Programming algorithm with Reinforcement learning presented in [1]. The new learning system allows for the utilisation of any pre-trained intelligent systems as processing nodes comprising the phenotypes. We envisage that compounding the GNP with more powerful processing nodes would extend its computing prowess. As proof of concept, we demonstrate that the extended evolutionary system can learn multi-behaviours for robots by testing it on the simulated Mirosot robot soccer domain to learn both target pursuit and wall avoidance behaviours simultaneously. A discussion of the development of the new evolutionary system is presented following an incremental order of complexity. The experiments show that the proposed algorithm converges to the desired multi-behaviour, and that the obtained system accuracy is better than a system that does not utilise pre-trained intelligent processing nodes.