Curiosity-driven exploration in reinforcement learning

Abstract

The paper elaborates upon a prior proposal for a novelty detector based on an artificial neural network forecaster. In the former paper, the novelty-based motivational signal was used in place of more conventional techniques (such as the ε-greedy policy, or the softmax policy) to drive exploration, in the context of V-learning. The current paper provides a more comprehensive study of such handling of the exploration vs. exploitation trade-off. It also studies the various problems concerning application of the approach to SARSA, and Q-learning. Also, and with the same goal in mind, the paper presents several advances upon the original design.