Show HN: A 178K Neural Net that beats Pokémon Roguelike (thiagolira.blot.im)

A developer has successfully created a reinforcement learning (RL) model capable of playing and winning a Pokémon roguelike game, with the model featuring only 178,000 parameters. The game entails navigating through nine maps filled with events that can involve capturing Pokémon or battling trainers, culminating in a final gauntlet against challenging opponents. The model, built using Proximal Policy Optimization (PPO), ingests a sparse vector encoding various states of the game, including the player's team composition and possible actions, enabling it to improve its gameplay through experience. This achievement is significant for the AI/ML community as it showcases how a relatively small neural network can effectively tackle complex decision-making tasks in gaming environments, where most approaches involve larger models. Key technical innovations include a structured approach to represent the action space concisely and the strategic use of dense rewards to guide learning, which ultimately allows the model to achieve a win rate of nearly 9%. This instance highlights the potential of reinforcement learning to solve real-world problems and the importance of carefully modeling game states and rewards to enhance performance.