Forward-Prediction-Based Active Exploration Behavior Cloning

Behavior cloning (BC) is a classical and effective imitation learning method for sequential decision problems in autonomous driving, robot control, and gameplay. However, behavior cloning suffers from the drawback of low sample utilization, which increases the interaction costs. In reinforcement learning applications, the interaction process for acquiring experience is often costly and risky. This study proposes a deep reinforcement learning algorithm called active exploration behavior cloning (AEBC) based on efficient exploration. Compared to the original behavior cloning algorithm, AEBC can achieve the desired performance with less data and imitate expert behavior more quickly. By integrating information entropy evaluation into the exploration term, the algorithm aims to enhance exploration. It is evaluated on multiple robot control tasks, demonstrating its performance and sampling efficiency. The experimental results show that the AEBC performs better in terms of sample efficiency than state-of-the-art behavior- cloning algorithms.