A Satisficing Strategy with Variable Reference in the Multi-armed Bandit Problems

The loosely symmetric model (LS) is as a subjective probability model that came from human beings' cognitive characteristics. To suggest a value to apply human beings' cognitive characteristics, we developed a value function "loosely symmetric model with variable reference" (LS-aVR) that expanded LS in the decision-amaking. It is important how get a reference value having an agent from environment to determine whether an algorithm using LS-aVR explores in comparison with a reference value. In this study, we proposed using statistical knowledge in an online method to acquire a reference value. Therefore we succeeded in making the result that new method exceeded a superior existing model in the multi-aarmed banded problem that is a kind of decision-amaking problems.