Positive-Unlabeled Reward Learning

Learning reward functions from data is a promising path towards achieving scalable Reinforcement Learning (RL) for robotics. However, a major challenge in training agents from learned reward models is that the agent can learn to exploit errors in the reward model to achieve high reward behaviors that do not correspond to the intended task. These reward delusions can lead to unintended and even dangerous behaviors. On the other hand, adversarial imitation learning frameworks (Ho & Ermon, 2016) tend to suffer the opposite problem, where the discriminator learns to trivially distinguish agent and expert behavior, resulting in reward models that produce low reward signal regardless of the input state. In this paper, we connect these two classes of reward learning methods to positive-unlabeled (PU) learning, and we show that by applying a large-scale PU learning algorithm to the reward learning problem, we can address both the reward under-and overestimation problems simultaneously. Our approach drastically improves both GAIL and supervised reward learning, without any additional assumptions. While Reinforcement Learning (RL) has shown itself to be a powerful tool for automating control and decision making, hand-specifying reward functions requires significant engineering effort, especially in real-world settings. Recent works have made promising progress in learning reward functions directly from human supervision, such as ratings (Cabi et al., 2019) and behavior preferences (Wilson et al., 2012; Ibarz et al., 2018). However, in practice, these supervisions are expensive to curate and thus often only cover a fraction of the state space. As a result, the learned reward functions may have large errors in the unlabeled states, and policy learning algorithms tend to exploit these errors to achieve extremely high pseudo-reward via unintended behaviors (Amodei et al., 2016). Practical solutions often require a human to provide supervision in the policy training loop iteratively (Christiano et al., 2017; Ibarz et al., 2018), resulting in a even more laborious process. On the other hand, works in Inverse Reinforcement Learning (IRL) propose to infer reward functions directly from expert behaviors (Ng et al., 2000; Ziebart et al., 2008), but scaling these methods to high-dimensional state space remains a challenge. Ho & Ermon (2016), and many followup works show that GAIL can learn complex behaviors even in high-dimensional spaces.