Reinforcement learning based dependency resolution

Описание

Dependency resolvers try to satisfy the version range requirements of libraries used using different approaches, such as backtracking or implementing an SAT solver. But what about plugging in machine learning algorithms in this AI/ML era and try to resolve software stacks with reinforcement learning? What are the benefits and drawbacks of such an approach? What is the state space imagined, what are the actions to be taken to accumulate rewards and how to balance exploration and exploitation in the state space? Let's play this game and see who wins...