Toward Agents That Reason About Their Computation
This addresses the issue of computational inefficiency in reinforcement learning agents, potentially leading to more energy-efficient AI systems, though it is incremental in applying existing methods to a new aspect of agent behavior.
The paper tackles the problem of reinforcement learning agents not becoming computationally efficient as they improve, by enabling agents to reason about and control their compute usage. The result shows that with the same training compute budget, these agents perform better on 75% of games and use three times less compute on average.
While reinforcement learning agents can achieve superhuman performance in many complex tasks, they typically do not become more computationally efficient as they improve. In contrast, humans gradually require less cognitive effort as they become more proficient at a task. If agents could reason about their compute as they learn, could they similarly reduce their computation footprint? If they could, we could have more energy efficient agents or free up compute cycles for other processes like planning. In this paper, we experiment with showing agents the cost of their computation and giving them the ability to control when they use compute. We conduct our experiments on the Arcade Learning Environment, and our results demonstrate that with the same training compute budget, agents that reason about their compute perform better on 75% of games. Furthermore, these agents use three times less compute on average. We analyze individual games and show where agents gain these efficiencies.