ASTEROID CHESS AND GO: REINFORCEMENT LEARNING FOR OPTIMAL SINGLE-INTERVENTION PLANETARY DEFENSE

Abstract In 1997, IBM's Deep Blue defeated chess world champion Garry Kasparov. In 2016, DeepMind's AlphaGo defeated Go world champion Lee Sedol 4–1 in a match experts predicted was a decade away. By 2017, AlphaGo Zero surpassed all prior versions within 40 days, learning entirely from self-play with no human data. In 2024, DeepMind's AlphaFold won the Nobel Prize in Chemistry for solving the protein folding problem — a challenge that had confounded biology for fifty years. In 2025, Apple's GIGAFLOW simulator trained autonomous driving policies across 1.6 billion kilometers of self-play — 9,500 years of driving experience in under 10 days on a single 8-GPU node — producing agents that outperform human drivers by a factor of four in safety without ever seeing human driving data. The trajectory is clear. Self-play reinforcement learning has progressed from board games to molecular biology to real-time physical navigation in nine years. This paper proposes the next domain: planetary defense. The solar system is a board. Asteroids and comets are pieces in motion. Orbital mechanics are the rules. The win condition is survival of all planets and moons. The available move is a single kinetic impulse — one DART-type mass-to-mass energy conversion. We argue that a reinforcement learning system trained through self-play against the full orbital dynamics of the solar system can identify optimal single-intervention strategies that no human team could compute, reducing planetary defense from reactive threat-by-threat monitoring to a solved optimization problem. Two physical frameworks for computing the energy transfer of the intervention are presented: the standard Newtonian momentum exchange, and the water-matrix lattice model in which gravitational coupling is frequency-dependent and the energy transfer involves lattice dynamics beyond simple momentum conservation.