Bayesian Approach to Learning Temporally Extended Concepts

In several previous studies it has been shown that the generalization capabilities of humans through concept learning is reminiscent of Bayesian modeling. When discriminating concepts from one another, human subjects tend to focus on the relevant features of the subspace and ignore the irrelevant ones. In this paper we propose a Bayesian concept learning paradigm that utilizes unrestricted Bayesian netw rks to learn the required concepts for optimal decision making. This approach has several beneficial characteristics that a concept learning algorithm should hold. At first it can both learn form observing an expert performing the desired task and from its own experience while carrying it out. Secondly, it is a close and computationally feasible approximation to the Bayesian modeling capabilities of humans. Thirdly, the Markov blanket surrounding the decision variable can render the irrelevant features independent and therefore this approach can ignore them seamlessly from the feature subspace. The simulation and experimental results are promising and show that our approach can successfully extract the required temporally extended concepts for a mobile robot task.

کلیدواژه ها:

Bayesian network ، Concept learning ، Imitation learning ، Reinforcement learning