Show HN: 178K Parameter Neural Net That Wins Poke(rogue)like (blog.thiagolira.com.br)

A developer has introduced a 178K parameter neural network that successfully navigates a Pokémon-inspired roguelike game using reinforcement learning techniques. This project highlights the effectiveness of a small model, trained using Proximal Policy Optimization (PPO), to tackle defined game mechanics such as states, actions, and rewards. By representing the game state as a 1386-dimensional sparse vector containing information about the player's team, current map, potential actions, and gym challenges, the model processes various game elements in a shared structure, optimizing its decisions for each state. This initiative is significant for the AI/ML community as it demonstrates how reinforcement learning can be applied to complex, interactive environments even with limited model capacity. The success of the neural network, which achieves a nearly 9% win rate, underscores the importance of designing dense rewards and providing comprehensive state visibility for effective learning. Moreover, the project serves as a reminder that clever structuring of input data and game mechanics can lead to successful outcomes, allowing simpler models to compete in challenging scenarios. This exploration contributes to the ongoing discourse on the capabilities of small-niche networks in overcoming hurdles traditionally reserved for larger, more complex architectures.