Earning with U Nsupervised a Uxiliary T Asks

Deep reinforcement learning agents have achieved state-of-the-art results by directly maximising cumulative reward. However, environments contain a much wider variety of possible training signals. In this paper, we introduce an agent that also learns separate policies for maximising many other pseudo-reward functions simultaneously by reinforcement learning. All of these tasks share a common representation that, like unsupervised learning, continues to develop in the absence of extrinsic rewards. We also introduce a novel mechanism for focusing this representation upon extrinsic rewards, so that learning can rapidly adapt to the most relevant aspects of the actual task. Our agent significantly outperforms the previous state-of-the-art on Atari, averaging 880% expert human performance, and a challenging suite of first-person, three-dimensional Labyrinth tasks leading to a mean speedup in learning of 10× and averaging 87% expert human performance on Labyrinth. Natural and artificial agents live in a stream of sensorimotor data. At each time step t, the agent receives observations ot and executes actions at. These actions influence the future course of the sensorimotor stream. In this paper we develop agents that learn to predict and control this stream, by solving a host of reinforcement learning problems, each focusing on a distinct feature of the sensorimotor stream. Our hypothesis is that an agent that can flexibly control its future experiences will also be able to achieve any goal with which it is presented, such as maximising its future rewards. The classic reinforcement learning paradigm focuses on the maximisation of extrinsic reward. However, in many interesting domains, extrinsic rewards are only rarely observed. This raises questions of what and how to learn in their absence. Even if extrinsic rewards are frequent, the sensorimotor stream contains an abundance of other possible learning targets. Traditionally, unsupervised learning attempts to reconstruct these targets, such as the pixels in the current or subsequent frame. It is typically used to accelerate the acquisition of a useful representation. In contrast, our learning objective is to predict and control features of the sensorimotor stream, by treating them as pseudorewards for reinforcement learning. Intuitively, this set of tasks is more closely matched with the agent’s long-term goals, potentially leading to more useful representations. Consider a baby that learns to maximise the cumulative amount of red that it observes. To correctly predict the optimal value, the baby must understand how to increase “redness” by various means, including manipulation (bringing a red object closer to the eyes); locomotion (moving in front of a red object); and communication (crying until the parents bring a red object). These behaviours are likely to recur for many other goals that the baby may subsequently encounter. No understanding of these behaviours is required to simply reconstruct the redness of current or subsequent images. Our architecture uses reinforcement learning to approximate both the optimal policy and optimal value function for many different pseudo-rewards. It also makes other auxiliary predictions that serve to focus the agent on important aspects of the task. These include the long-term goal of predicting cumulative extrinsic reward as well as short-term predictions of extrinsic reward. To learn more efficiently, our agents use an experience replay mechanism to provide additional updates ∗Joint first authors. Ordered alphabetically by first name.