Avoiding Negative Side-Effects and Promoting Safe Exploration with Imaginative Planning

Dhruv Ramani, Benjamin Eysenbach

Keywords: Reinforcement Learning, AI-Safety, Model-Based Reinforcement Learning, Safe-Exploration

Abstract: With the recent proliferation of the usage of reinforcement learning (RL) agents for solving real-world tasks, safety emerges as a necessary ingredient for their successful application. In this paper, we focus on ensuring the safety of the agent while making sure that the agent does not cause any unnecessary disruptions to its environment. The current approaches to this problem, such as manually constraining the agent or adding a safety penalty to the reward function, can introduce bad incentives. In complex domains, these approaches are simply intractable, as they require knowing apriori all the possible unsafe scenarios an agent could encounter. We propose a model-based approach to safety that allows the agent to look into the future and be aware of the future consequences of its actions. We learn the transition dynamics of the environment and generate a directed graph called the imaginative module. This graph encapsulates all possible trajectories that can be followed by the agent, allowing the agent to efficiently traverse through the imagined environment without ever taking any action in reality. A baseline state, which can either represent a safe or an unsafe state (based on whichever is easier to define) is taken as a human input, and the imaginative module is used to predict whether the current actions of the agent can cause it to end up in dangerous states in the future. Our imaginative module can be seen as a ``plug-and-play'' approach to ensuring safety, as it is compatible with any existing RL algorithm and any task with discrete action space. Our method induces the agent to act safely while learning to solve the task. We experimentally validate our proposal on two gridworld environments and a self-driving car simulator, demonstrating that our approach to safety visits unsafe states significantly less frequently than a baseline.