Researchers from Sakana AI and MIT developed a new strategy for the Core War programming game using evolutionary processes and adversarial dynamics. The approach utilizes the MAP-Elites optimization algorithm to find a diverse set of high-performing programs, rather than a single ideal solution.
The system employs GPT-4.1 mini (gpt-4.1-mini-2025-04-14) to generate new warriors from scratch and provide intelligent improvements to existing code. This replaces the traditional method of random symbol mutation in the Redcode language.
Testing involved 96 training opponents and 317 test opponents, all created by humans. The researchers used behavioral descriptors such as the total number of spawned threads and the total memory volume of the warrior during simulation to analyze performance.
The results showed that warriors became more universal and resilient against unseen strategies as the number of epochs increased. The authors concluded that the DRQ method encourages similar behavior among warriors while maintaining diversity in their basic implementations, mirroring convergent evolution in biology.
To predict the outcome for a warrior, the authors applied linear regression using OpenAI text-embedding-3-small and OpenAI text-embedding-3-large. This demonstrated a moderate predictive power, indicating that while the code partially determines effectiveness, the battle dynamics also play a significant role.