抄録
We have proposed an on-line adaptive system for Improving the reinforcement learning (RL) agents' Policies by using a Mixture of Bayesian Networks (IPMBN).
Results of computer simulations and robotic experiments revealed that RL agents or robots with IPMBN could adapt to a variety of environmental variation by means of policy (or environment) information represented by the BN mixtures.
In both the simulations and the experiments, the data for structure learning of BNs were generated and collected on ``idealized'' virtual spaces in which any noise, which can exist in real-world environments, was strictly removed.
Therefore, we have not yet verify whether the data (i.e. experience) of learner generated in real-world environments containing noise is available for the learner's adaptaion to changes in the real world environment without modification.
In this article, some kind of noise are added to input of an RL agent behaving in virtual environments in order to evaluate the ability of representing policy information in the mixture of BNs, each structure of which is decided by means of the data collected in such environments.
We also discuss the adaptability of the agents with our proposed system using the above-mentioned BN mixtures to the environmental changes.
