Optimizing Hybrid Multi-Task Deep Reinforcement Learning: Diverse Perspectives

This study explores Deep Reinforcement Learning (DRL) by integrating Reinforcement Learning (RL) into a lightweight framework using Deep Learning (DL). Despite DRL's high performance, challenges include data hunger, extensive computation, and prolonged training times. To address these, this paper proposes a multi-task learning approach for optimizing DRL agents. The focus is on the Hybrid Multi-Task Asynchronous Advantage Actor-Critic (A3C) algorithm, demonstrating its stabilizing effect on training across diverse video games. The implementation uses Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) in Actor and Critic networks, showcasing significant enhancements in specific game environments. In spite of training in two environments concurrently, the average time required is still comparable to the A3C training in a single environment.