Incremental learning approach for board game playing agents

摘要

In this paper an incremental learning method for neural network-based agent is introduced. The underlying idea of the presented method is the effective use of previously acquired knowledge in subsequent learning. Consequently, after a certain period of training, the agent gains the ability to gradually improve the learning process. The improvement is achieved due to the mechanism of sharing relevant features among "similar" learning problems along with a gradual increase of the complication level of the training examples. Preliminary experimental results are presented for the Othello game playing agent.