Learning state-based behaviour using temporally related cases

摘要

Learning by observation allows a software agent to learn an expert’s behaviour, by examining the actions the expert performs in response to inputs, without the expert having to explicitly program the agent. Most learning by observation approaches only make use of the current inputs and actions of the expert and ignore any past inputs or actions. This limits the agents to only being able to learn reactive behaviour. We present an approach to case retrieval that uses the expert’s past inputs and actions in order to allow for learning state-based behaviour. We demonstrate our approach by learning from a simulated obstacle avoidance robot that reasons using internal state information. Our results show a significant accuracy improvement over retrieval that does not take into account any past information.