Attention-based Curiosity-driven Exploration in Deep Reinforcement Learning

TL;DR

This paper introduces attention mechanisms into curiosity-driven exploration methods in deep reinforcement learning, proposing new models like AttA2C and enhancements to ICM to improve exploration efficiency in sparse reward environments.

Contribution

It extends curiosity-driven exploration by integrating attention mechanisms and proposes novel methods such as AttA2C and rational curiosity, enhancing exploration in deep RL.

Findings

AttA2C improves exploration efficiency.

Attention-based ICM emphasizes relevant features.

Rational curiosity offers a new intrinsic reward formulation.

Abstract

Reinforcement Learning enables to train an agent via interaction with the environment. However, in the majority of real-world scenarios, the extrinsic feedback is sparse or not sufficient, thus intrinsic reward formulations are needed to successfully train the agent. This work investigates and extends the paradigm of curiosity-driven exploration. First, a probabilistic approach is taken to exploit the advantages of the attention mechanism, which is successfully applied in other domains of Deep Learning. Combining them, we propose new methods, such as AttA2C, an extension of the Actor-Critic framework. Second, another curiosity-based approach - ICM - is extended. The proposed model utilizes attention to emphasize features for the dynamic models within ICM, moreover, we also modify the loss function, resulting in a new curiosity formulation, which we call rational curiosity. The corresponding implementation can be found at https://github.com/rpatrik96/AttA2C/.