Adopting animal concepts in hierarchical reinforcement learning and control of intelligent agents
David Kadleček, Pavel Nahodil · 2008
This research integrates rigorous methods of reinforcement learning (RL) and control engineering with a behavioral (ethology) approach to the agent technology. The main outcome is a hybrid architecture for intelligent autonomous agents targeted to the Artificial Life like environments. The architecture adopts several biology concepts and shows that they can provide robust solutions to some areas. The resulting agents perform from primitive behaviors, simple goal directed behaviors, to complex planning. The agents are fully autonomous through environment feedback evaluating internal agent state and motivates the agent to perform behaviors that return the agent towards optimal conditions. This principle is typical to animals. Learning and control is realized by multiple RL controllers working in a hierarchy of Semi Markov Decision Processes (SMDP). Used model free Q(lambda) learning works online, the agents gain experiences during interaction with the environment. The decomposition of the root SMDP into hierarchy is automated as opposed to the conventional methods that are manual. The agents assess utility of the behavior and provide rewards to RL controller as opposed to the conventional RL methods where the rewards-situations map is defined by the designer upfront. The resulting learning algorithm converges to a recursively optimal solution with probability 1. Agent behavior is continuously optimized according to the distance from the agentpsilas optimal conditions.