Deep neural networks have achieved impressive successes in fields ranging from object recognition to complex games such as Go. Navigation, however, remains a substantial challenge for artificial agents, with deep neural networks trained by reinforcement learning failing to rival the proficiency of mammalian spatial behaviour, which is underpinned by grid cells in the entorhinal cortex . Grid cells are thought to provide a multi-scale periodic representation that functions as a metric for coding space and is critical for integrating self-motion (path integration) and planning direct trajectories to goals (vector-based navigation). Here we set out to leverage the computational functions of grid cells to develop a deep reinforcement learning agent with mammal-like navigational abilities. We first trained a recurrent network to perform path integration, leading to the emergence of representations resembling grid cells, as well as other entorhinal cell types . We then showed that this representation provided an effective basis for an agent to locate goals in challenging, unfamiliar, and changeable environments-optimizing the primary objective of navigation through deep reinforcement learning. The performance of agents endowed with grid-like representations surpassed that of an expert human and comparison agents, with the metric quantities necessary for vector-based navigation derived from grid-like units within the network. Furthermore, grid-like representations enabled agents to conduct shortcut behaviours reminiscent of those performed by mammals. Our findings show that emergent grid-like representations furnish agents with a Euclidean spatial metric and associated vector operations, providing a foundation for proficient navigation. As such, our results support neuroscientific theories that see grid cells as critical for vector-based navigation, demonstrating that the latter can be combined with path-based strategies to support navigation in challenging environments.

No abstract

The reinforcement learning community has made great strides in designing algorithms capable of exceeding human performance on specific tasks. These algorithms are mostly trained one task at the time, each new task requiring to train a brand new agent instance. This means the learning algorithm is general, but each solution is not; each agent can only solve the one task it was trained on. In this work, we study the problem of learning to master not one but multiple sequentialdecision tasks at once. A general issue in multi-task learning is that a balance must be found between the needs of multiple tasks competing for the limited resources of a single learning system. Many learning algorithms can get distracted by certain tasks in the set of tasks to solve. Such tasks appear more salient to the learning process, for instance because of the density or magnitude of the in-task rewards. This causes the algorithm to focus on those salient tasks at the expense of generality. We propose to automatically adapt the contribution of each task to the agent's updates, so that all tasks have a similar impact on the learning dynamics. This resulted in state of the art performance on learning to play all games in a set of 57 diverse Atari games. Excitingly, our method learned a single trained policy -with a single set of weights -that exceeds median human performance. To our knowledge, this was the first time a single agent surpassed human-level performance on this multi-task domain. The same approach also demonstrated state of the art performance on a set of 30 tasks in the 3D reinforcement learning platform DeepMind Lab.

1Over the past twenty years, neuroscience research on reward-based learning has converged on a canonical model, under which the neurotransmitter dopamine 'stamps in' associations between situations, actions and rewards by modulating the strength of synaptic connections between neurons. However, a growing number of recent findings have placed this standard model under strain. In the present work, we draw on recent advances in artificial intelligence to introduce a new theory of rewardbased learning. Here, the dopamine system trains another part of the brain, the prefrontal cortex, to operate as its own free-standing learning system. This new perspective accommodates the findings that motivated the standard model, but also deals gracefully with a wider range of observations, providing a fresh foundation for future research.Exhilarating advances have recently been made toward understanding the mechanisms involved in reward-driven learning. This progress has been enabled in part by the importation of ideas from the field of reinforcement learning 1 (RL). Most centrally, this input has led to an RL-based theory of dopaminergic function. Here, phasic dopamine (DA) release is interpreted as conveying a reward prediction error (RPE) signal [2][3][4] , an index of surprise which figures centrally in temporal-difference RL algorithms 1 . Under the theory, the RPE drives synaptic plasticity in the striatum, translating experienced action-reward associations into optimized behavioral policies 4,5 . Over the past two decades, evidence has steadily mounted for this proposal, establishing it as the standard model of reward-driven learning.However, even as this standard model has solidified, a collection of problematic observations has accumulated. One quandary arises from research on prefrontal cortex (PFC). A growing body of evidence suggests that PFC implements mechanisms for reward-based learning, performing computations that strikingly resemble those ascribed to DA-based RL. It has long been established that sectors of the PFC represent the expected values of actions, objects and states [6][7][8] . More recently, it has emerged that PFC also encodes the recent history of actions and rewards [9][10][11][12][13][14][15] . The set of variables encoded, along with observations concerning the temporal profile of neural activation in the PFC, has led to the conclusion that "PFC neurons dynamically [encode] conversions from reward and choice history to object value, and from object value to object choice" 10 . In short, neural activity in PFC appears to reflect a set of operations that together constitute a self-contained RL algorithm.Placing PFC beside DA, we obtain a picture containing two full-fledged RL systems, one utilizing activity-based representations and the other synaptic learning. What is the relationship between these systems? If both support RL, are their functions simply redundant? One suggestion has been that DA and PFC subserve different forms of learning, with DA implementing model-free RL, based on direct stim...

In this work we aim to solve a large collection of tasks using a single reinforcement learning agent with a single set of parameters. A key challenge is to handle the increased amount of data and extended training time. We have developed a new distributed agent IMPALA (Importance Weighted Actor-Learner Architecture) that not only uses resources more efficiently in singlemachine training but also scales to thousands of machines without sacrificing data efficiency or resource utilisation. We achieve stable learning at high throughput by combining decoupled acting and learning with a novel off-policy correction method called V-trace. We demonstrate the effectiveness of IMPALA for multi-task reinforcement learning on DMLab-30 (a set of 30 tasks from the DeepMind Lab environment (Beattie et al., 2016)) and Atari-57 (all available Atari games in Arcade Learning Environment (Bellemare et al., 2013a)). Our results show that IMPALA is able to achieve better performance than previous agents with less data, and crucially exhibits positive transfer between tasks as a result of its multi-task approach. The source code is publicly available at github.com/deepmind/scalable agent.