Enhanced deep reinforcement learning deep q-network models

Abstract

A reinforcement learning method and apparatus includes storing video frames in a video memory, performing a first preprocessing step of retrieving a sequence of n image frames of the stored video frames, and merging the n image frames in a fading-in fashion by incrementally increasing the intensity of each frame up to the most recent frame having full intensity to obtain a merged frame; and performing a training step of inputting the merged frame to the DQN and training the DQN to learn Q-values for all possible actions from a state represented by the merged frame with only a single forward pass through the network. The learning method and apparatus includes a second preprocessing step of removing the background from the merged frame. The method can be applied to any DQN learning method that uses a convolution neural network as its core value function approximator.