Reinforcement Learning for Source Location Estimation: A Multi-Step Approach

Gas leaks present an undeniable safety concern; the ability to swiftly and accurately detect the source of a leak and pertinent details is critical for effective emergency response. The limited precision of sensors and environmental noise introduce significant uncertainty and randomness, complicating the resolution of such issues. To address these challenges, this study introduces a new approach that integrates multi-step deep reinforcement learning algorithms with Bayesian inference to estimate source information. Compared to single-step Reinforcement Learning and Entrotaxis methods, the multi-step update mechanism in this problem allows the agent to locate sources position more efficiently. This approach not only increases the search's success rate but also decreases the number of time steps needed for successful detection. Experiments conducted in continuous and discrete environments of equal scale and parameters corroborate the efficiency of our method in tracing the source of gas leaks.