Researcher Mikhail Kocherov is developing a reinforcement learning (RL) agent to master the Atari Frostbite game. The project aims to explore brain mechanisms by simulating autonomous intelligent behavior, where the agent learns to navigate environments, manage memory, and optimize actions based on rewards.
The agent operates within the Atari Learning Environment (ALE) and Gymnasium. Frostbite presents challenges such as non-stationary game dynamics, sparse rewards, and hierarchical goal setting. The agent is tasked with tasks like jumping on ice floes, avoiding enemies, and building igloos, with the goal of surpassing the current state-of-the-art MEME model.
The project utilizes the Proximal Policy Optimization (PPO) algorithm, which is chosen for its scalability and efficiency in parallelized environments. The architecture includes an encoder for processing 210x160 RGB frames, an action head for decision-making, and a value head for predicting rewards. The system processes trajectories to refine its policy through iterative optimization.
The training process involves a battery of 32 simultaneous game instances. The agent learns by sampling actions from a probability distribution, where successful moves are reinforced over time. The model compresses raw visual input into 384-dimensional latent vectors, balancing memory constraints with the need to interpret temporal changes in the game environment.
