Never Give Up: Learning Directed Exploration Strategies

By Adrià Puigdomènech Badia, Pablo Sprechmann, Alex Vitvitskyi, Daniel Guo, Bilal Piot, Steven Kapturowski, Olivier Tieleman, Martín Arjovsky, Alexander Pritzel, A. N. Bolt and 1 more. 2020

360 citations recorded by the source corpus.

We propose a reinforcement learning agent to solve hard exploration games by learning a range of directed exploratory policies. We construct an episodic memory-based intrinsic reward using k-nearest neighbors over the agent's recent experience to train the directed exploratory policies, thereby encouraging the agent to repeatedly revisit all states in its environment. A self-supervised inverse dynamics model is used to train the embeddings of the nearest neighbour lookup, biasing the novelty signal towards what the agent can control. We employ the framework of Universal Value Function Approximators (UVFA) to simultaneously learn many directed exploration policies with the same neural network, with different trade-offs between exploration and exploitation. By using the same neural network for different degrees of exploration/exploitation, transfer is demonstrated from predominantly exploratory policies yielding effective exploitative policies. The proposed method can be incorporated to run with modern distributed RL agents that collect large amounts of experience from many actors running in parallel on separate environment instances. Our method doubles the performance of the base agent in all hard exploration in the Atari-57 suite while maintaining a very high score across the remaining games, obtaining a median human normalised score of 1344.0%. Notably, the proposed method is the first algorithm to achieve non-zero rewards (with a mean score of 8,400) in the game of Pitfall! without using demonstrations or hand-crafted features.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Vision and Pattern Recognition

Citations in the corpus

Acme: A Research Framework for Distributed Reinforcement Learning (2020, background): “Examples include modifications to the reward signals to include intrinsic rewards (Bellemare et al., 2016; Badia et al., 2020) or auxiliary tasks (Jaderberg et al.”
A Kernel-Based Approach to Non-Stationary Reinforcement Learning in Metric Spaces (2020, background): “In larger dimension, kernel-based exploration bonuses have been recently shown to enhance exploration in RL for Atari games (Badia et al., 2020), and our approach might inspire the de-sign of bonuses for high-dimensional non-stationary environments.”
Rapid Task-Solving in Novel Environments (2020, methodology): “Recently, promising methods for exploration with deep-RL have been introduced, including Go-Explore (Ecoffet et al., 2019) and Never Give Up (Badia et al., 2020).”
Agent57: Outperforming the Atari Human Benchmark (2020, methodology): “The reason for choosing this different loss is that it worked better than the n-step loss for NGU, as described in Puigdomènech Badia et al. (2020).”
Reactive Exploration to Cope with Non-Stationarity in Lifelong Reinforcement Learning (2022, methodology): “Other strategies include memory-based exploration strategies such as episodic memory (Badia et al., 2020) or planning with worst-base planning evolution (Lecarpentier & Rachelson, 2019), GoExplore (Ecoffet et al.”
Semantic Exploration from Language Abstractions and Pretrained Representations (2022, methodology): “Exploration in RL Classical exploration strategies include (cid:15) -greedy action selection [51], state-counting [49, 4, 32, 30, 3], curiosity driven exploration [44], and intrinsic motivation methods [36].”
Bootstrap Latent-Predictive Representations for Multitask Reinforcement Learning (2020, background): “Representation learning has also had broader impact in deep RL, as representations have been used to generate intrinsic motivation signal for exploration (Houthooft et al., 2016; Pathak et al., 2017; Burda et al., 2019; Azar et al., 2019; Puigdomènech Badia et al., 2020) and as the basis for search ”
Multi-Stage Episodic Control for Strategic Exploration in Text Games (2022, neutral): “Other works learn a mixture of policies for decoupling exploration and exploitation, either by using a conditional architecture with shared weights (Badia et al., 2020), pre-deﬁning an exploration mechanism for restricted policy optimization (Shani et al., 2019), or learning separate task and…”
BeBold: Exploration Beyond the Boundary of Explored Regions (2020, background): “Various criteria have been proposed, including curiosity/surprise -driven (Pathak et al., 2017), count -based (Bellemare et al., 2016; Burda et al., 2018a;b; Ostrovski et al., 2017; Badia et al., 2020b), and state-diff approaches (Zhang et al., 2019; Marino et al., 2019).”