Generalised Discount Functions applied to a Monte-Carlo AImu Implementation
This work provides a practical demonstration of theoretical discounting concepts in GRL, but it is incremental as it applies known methods to a new simulation context without broad impact.
The authors tackled the lack of concrete demonstrations of generalized discounting in General Reinforcement Learning by implementing arbitrary discount functions in the AIXIjs platform and testing them in a simple MDP. They found that agent behavior aligned with theoretical expectations when appropriate Monte-Carlo Tree Search parameters were used, but did not report specific numerical results.
In recent years, work has been done to develop the theory of General Reinforcement Learning (GRL). However, there are few examples demonstrating these results in a concrete way. In particular, there are no examples demonstrating the known results regarding gener- alised discounting. We have added to the GRL simulation platform AIXIjs the functionality to assign an agent arbitrary discount functions, and an environment which can be used to determine the effect of discounting on an agent's policy. Using this, we investigate how geometric, hyperbolic and power discounting affect an informed agent in a simple MDP. We experimentally reproduce a number of theoretical results, and discuss some related subtleties. It was found that the agent's behaviour followed what is expected theoretically, assuming appropriate parameters were chosen for the Monte-Carlo Tree Search (MCTS) planning algorithm.