Shallow trees with heavy leaves (2020) (cp4space.hatsya.com)

A new computational search program called ikpx2 has been developed, inspired by the principles of deep learning used in Leela Chess Zero, a descendant of DeepMind’s AlphaZero. Unlike Stockfish, which evaluates 100 million positions per second using traditional heuristics, ikpx2 explores a smaller search tree but puts more computational effort into assessing each position. It utilizes a SAT solver to perform deep lookahead, enabling it to predict whether a partial spaceship in cellular automata can continue evolving, thereby optimizing resource use. This approach allows for significant memory efficiency and enables parallel processing, making it easier to manage the search tasks. The significance of ikpx2 lies in its innovative method of integrating self-play and reinforcement learning to enhance its SAT solver selection based on the specific task at hand. This strategy not only leads to faster search results—achieving a notable 50-fold increase in efficiency compared to its predecessor—but also expands the exploration into various cellular automata rules by encoding complex mechanics into SAT problems. The program's open-source nature invites further experimentation and development within the AI/ML community, paving the way for enhanced discoveries in both computational algorithms and the study of cellular automata.