This project aims to investigate the generalization abilities of various reinforcement learning (RL) algorithms across a set of previously untested tasks. Specifically, we will examine the performance of an ExpGen inspired algorithm and compare it with the classical Proximal Policy Optimization (PPO) algorithm. The chosen benchmark environment for this task is Crafter, an open-world survival game with visual inputs that evaluates a wide range of general abilities within a single environment.
We conducted a thorough review of the literature to identify state-of-the-art (SOTA) algorithm that performs well on this environment. Through this analysis, we gained valuable insights into the key factors contributing to its success in Crafter. Drawing inspiration from these insights, we designed an inspired ExpGen algorithm. We then conducted a comparative evaluation of PPO, our ExpGen-inspired algorithm, and SOTA-inspired variants of the ExpGen approach, to assess their relative strengths and generalization abilities on Crafter.