Abstract
Penalty Avoiding Rational Policy Making algorithm (PARP) based on Profit Sharing method and was planed to learn a penalty avoiding policy. PARP is improved to save memories and to cope with uncertainties. The efficiency of the Improved Penalty Avoiding Rational Policy Making algorithm is influenced by threshold of the penalty basis function γ significantly. Up to now, it is necessary to set appropriate γ through a preliminary experiment. In this paper, we propose a technique for learning γ with the multi start method. The proposal technique is applied to a keepaway task that is a benchmark in a robotic soccer game, to confirm the effectiveness.
