Learning to select computations

Efficient use of limited computational resources is essential to intelligence. Selecting computations optimally according to rational metareasoning would achieve this, but rational metareasoning is computationally intractable. Inspired by psychology and neuroscience, we propose the first learning algorithm for approximating the optimal selection of computations. We derive a general, sample-efficient reinforcement learning algorithm for learning to select computations from the insight that the value of computation lies between the myopic value of computation and the value of perfect information. We evaluate the performance of our method against two state-of-theart methods for approximate metareasoning–the meta-greedy heuristic and the blinkered policy–on three increasingly difficult metareasoning problems: metareasoning about when to terminate computation, metareasoning about how to choose between multiple actions, and metareasoning about planning. Across all three domains, our method achieved near-optimal performance and significantly outperformed the meta-greedy heuristic. The blinkered policy performed on par with our method in metareasoning about decision-making, but it is not directly applicable to metareasoning about planning where our method outperformed both the meta-greedy heuristic and a generalization of the blinkered policy. Our results are a step towards building self-improving AI systems that can learn to make optimal use of their limited computational resources to efficiently solve complex problems in real-time.