💊Strategy Builder

The Strategy Builder platform offers an innovative approach to AI research through a game-like interface. Instead of directly playing the game, participants craft strategies and directives that guide AI-controlled characters in simulated combat scenarios. These simulations are then rendered and presented back to users, creating an engaging feedback loop. This unique system provides an entertaining experience for users, generates valuable data for AI research, and allows participants to contribute to scientific advancements without requiring specialised skills. By creating prompts and strategies, users help develop and refine AI models' understanding of complex, real-time decision-making processes. This system works by using a Large Language Model (LLM) to control the actions of a player in a game environment.

The LLM receives a text-based description of the game screen, including details like the player's previous moves, the opponents' moves, and the status of power and health bars. Based on this information, the LLM determines the next move for its character. The flow involves multiple inputs such as game observations, relative positions, possible actions, and reward functions, which are fed into a system prompt. Each player in the game is controlled by an LLM observing game states, calculating rewards, and determining possible actions. It’s important to note that LLMs are inherently non-deterministic, meaning that even when given the exact same input, they can produce different outputs due to the randomness introduced during the generation process.