Analyzing State Space Similarities for Multi-Task Deep Reinforcement Learning in Atari Games

Deep reinforcement learning (DRL) has shown effectiveness across various individual tasks, but often faces challenges when applied to multitask scenarios. This study investigates the potential for sequential multitask DRL by analyzing the similarities within the state spaces of different Atari game environments. Using dynamic time warping and dimensionality reduction techniques, we extract meaningful features from the state representations and compare their similarities. The result provides insights into the relationships between distinct Atari games. This work aims to establish a foundation for future research focused on developing more efficient and robust multitask DRL algorithms that can leverage shared knowledge across similar tasks.