Abstract
Recently, many works about Partially Observable Markov Decision Processes have attracted increasing attentions in reinforcement learning because the world cannot be always modeled on Markov Decision Processes. In this paper, Time-dependent Classifier System; TCS is proposed to solve perceptual aliasing problems. In TCS, it keeps the constant length of time sequential information, and uses them only when it is necessary to solve the perceptual aliasing. So, the changeable lengths of rules are proposed in TCS using the time tags. And the internal action to detect the information of 1 time-step is used to link the present information to the past information. And the several maze problems with Partially Observable Markov Decision Processes are experimented to verify TCS's effectiveness.
