S-learning: a reinforcement learning method without parameter tuning

Recently, literature reported progresses on two reinforcement learning algorithms, AHC [1] and Q-learning [2]. We discussed issues raised in these conventional formulations from the aspects of convergence, parameter-tuning, over-training, computational and storage efficiency, and then proposed two new reinforcement learning mechanisms: S-learning and S&AHC learning. Particularly, the representation of the final cost map formed by S-learning series can be explicitly interpreted as the number of minimum movements to the goal state from each individual state. In addition, an adaptive (incremental) S-learning was proposed which incorporated S-learning and the technique of incremental learning [3] to facilitate the practical implementation of neural reinforcement learning. All of S-learning series showed promising performances in exploring Sutton's task [4] of navigating in a maze.