A production technique for a Q-table with an influence map for speeding up Q-learning

Q-learning is a reinforcement learning widely used for automatic learning in the game environment. Before applying Q-learning, the many states of environment that an agent may come in contact with is defined. The weak point of Q-learning is the time it takes to learn these states as states become larger. In this paper, the Q-learning mechanism using an influence map (QIM) is proposed to reduce the time needed for learning. By using an influence map and the learning result, a medium Q-value, which is not yet learnt, will be generated. Generally, when learning is finished, it is difficult to improve the performances. If QIM is used, however, the performance could be improved. Although the Q-table in QIM has been defined with small states, QIM obtains nearly the same learning result.