Point-Based Value Iteration — LessWrong