Autoresearch for SAT Solvers (github.com)

An innovative AI agent has been developed to autonomously become a leading expert in MaxSAT, a challenging optimization problem, by training on 229 instances from the 2024 MaxSAT Evaluation without any human intervention. This agent leverages a collaborative GitHub framework where each instance of the agent can share knowledge and improvements in real-time. It utilizes advanced techniques like Greedy SAT, core-guided search, and dynamic clause weighting, discovering effective strategies and achieving notable breakthroughs, including solving instances previously deemed unsolvable. This development is significant for the AI/ML community as it demonstrates the potential for self-improving systems in complex problem domains. The agent solved 220 out of 229 instances and outperformed existing competition solutions in several cases, showcasing not only superior optimization capabilities but also the ability to innovate. Key technical implications include the agent's ability to refine its approach through experimentation, as seen in various techniques that enhance performance, such as biased SAT for breaking local optima and alternating search strategies for deep optimization. This project exemplifies how autonomous learning can push the boundaries of AI research and applications in combinatorial optimization.