Online learning of fuzzy behaviour co-ordination for autonomous agents using genetic algorithms and real-time interaction with the environment

Addresses the development of a system for online learning of fuzzy behaviour co-ordination for autonomous agents in the form of robots based on genetic algorithms (GAs) and real-time interaction with the environment. The proposed system organises the behaviours hierarchically and uses fuzzy engines to implement both the behaviours and their co-ordination mechanism. In previous work (1999) we reported on our success in the online learning of individual behaviours (rules and membership functions). In this paper we report on a system that allows the fuzzy membership function (MF) for behaviour co-ordination to be learnt online in a manner that satisfies some high level mission or plan. The GAs use adaptive learning parameters and guided constrained optimisation to speed the GAs search and enable it to be performed via real-world interaction rather than off-line simulation. The results of this work are compared with results reported elsewhere and reveals this approach to have a superior learning performance while learning using real outdoor robots in changing environments. The ability to learn co-ordination skills in a short time interval without human intervention makes this approach particularly useful for applications where access is difficult such as nuclear reactors, underwater vehicles and space robots and fast changing and dynamic environments such as the agricultural environments.