Behaviour Selection on a Mobile Robot using W-learning
Martin Hallerdal, John C. T. Hallam · The MIT Press eBooks · 2002
A common approach in complex reinforcement learning tasks is to divide the problem into functional parts, or behaviours, and then to assign a sub-agent to solve each task. The action selection problem then becomes to negotiate between sub-agents with conflicting desires. W-learning is a method whereby agents build up W-values in each state that indicate how important that state is for that agent. These values are then used as basis for selecting agents. In this paper we present the first results, as far as we know, of applying W-learning on a mobile robot in solving a task in the real world. Results from the experiments are presented and the suitability of W-learning for real world robot tasks is discussed.