Reward is (almost) enough
Keywords: Rewards, RL, intelligence, humans, nonstationarity, reframing
Abstract: For centuries, humans have sought to understand intelligence and its associated mechanisms that drive how we think. While some have hypothesized that distinct signals or objectives are required for different types of abilities including learning, perception, social intelligence, generalization and imitation, others have suggested that learning through trial and error to maximise rewards can help develop behaviours that encompass all of these abilities. In this paper, we posit that while maximising rewards is central to developing a diverse range of abilities, the way in which we think about and formulate these rewards has to be re-framed as the conventional approach to using rewards in reinforcement learning can be prohibitive and is known to underperform in various settings, including sparse environments and noisy reward conditions. We suggest that these rewards need to be reformulated to incorporate different notions of i) uncertainty, ii) human preferences, and iii) nested or mixed compositions, iv) non-stationarity as well as account for v) situations where no reward is necessary. We suggest that doing so could enable more powerful reinforcement learning agents as a step towards artificial general intelligence.
Submission Number: 38
Loading