Graphical modeling for dynamic safety hints generalisation for Safe Deep Reinforcement Learning Agents

Lamogha Chiazor, Lan Hoang, Shivam Ratnakar, Minta Shiharath

Keywords: safe reinforcement learning, graphical modelling

TL;DR: We propose to dynamically generate safety hints for DRL agents in a text based game environment to introduce external system dynamic knowledge to a safety-relevant agent which help improve the safety performance.

Abstract: A major challenge for Deep Reinforcement Learning (DRL) research is to maintain robustness and safety, which is highly important in real-life applications.There is a need to explore AI models/algorithms to utilise system dynamics and their associated reward distribution (assessing risk/opportunities trade-off) to develop algorithms for data-driven Markov Decision Process formulation and safe action sets. Causality modelling is a well-established logical technique which can help identify the effects of interactions of complex evolving systems involving several entities. In this paper, we propose using knowledge graphs which help DRL agents incorporate safety aspects of actions and entities in their decision making process. We dynamically generate safety hints for DRL agents in a text based game environment using a safety concept-net. Our experiments show that DRL agents with safety-hints perform better on safety-based games than agents without them.