Generalizing from a few environments in safety-critical reinforcement learning
arXiv:1907.01475
Abstract
Before deploying autonomous agents in the real world, we need to be confident they will perform safely in novel situations. Ideally, we would expose agents to a very wide range of situations during training, allowing them to learn about every possible danger, but this is often impractical. This paper investigates safety and generalization from a limited number of training environments in deep reinforcement learning (RL). We find RL algorithms can fail dangerously on unseen test environments even when performing perfectly on training environments. Firstly, in a gridworld setting, we show that catastrophes can be significantly reduced with simple modifications, including ensemble model averaging and the use of a blocking classifier. In the more challenging CoinRun environment we find similar methods do not significantly reduce catastrophes. However, we do find that the uncertainty information from the ensemble is useful for predicting whether a catastrophe will occur within a few steps and hence whether human intervention should be requested.
References in corpus (12)
- Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks
- Theoretical Models of Learning to Learn
- IMPALA: Scalable Distributed Deep-RL with Importance Weighted Actor-Learner Architectures
- RL: Fast Reinforcement Learning via Slow Reinforcement Learning
- Learning to reinforcement learn
- A Study on Overfitting in Deep Reinforcement Learning
- Uncertainty-Aware Reinforcement Learning for Collision Avoidance
- Quantifying Generalization in Reinforcement Learning
- AI Safety Gridworlds
- Generalization and Regularization in DQN
- Evaluating Uncertainty Quantification in End-to-End Autonomous Driving Control
- Combating Reinforcement Learning's Sisyphean Curse with Intrinsic Fear