A meta-heuristic algorithm combined with deep reinforcement learning for multi-sensor positioning layout problem in complex environment

Abstract

In a multi-sensor positioning system (MSPS), the layout of sensors plays a crucial role in determining the system’s performance. Therefore, addressing the sensor layout problem (SLP) within the MSPS is an essential approach to achieve high-precision location information. However, equipment failures and measurement losses in complex working conditions can disrupt the established sensor layout geometry, resulting in significant degradation of positioning accuracy. To address this issue, we introduce robustness as a new objective for sensor layout optimization within MSPS operating in complex environments, transforming it into a constrained multi-objective optimization problem. Consequently, we propose a Constrained Pareto Dominance Evolutionary Algorithm based on Deep Q Network (CDEA-DQN). This algorithm incorporates a state quaternion that characterizes population quality in both objective and decision spaces. It further establishes a mapping model from state to optimal reproduction operators while employing reward and update strategies that provide adaptive preferences for convergence, diversity, and feasibility – enabling dynamic reproduction. Experimental results from 44 benchmark instances along with three proposed SLP scenarios demonstrate the effectiveness of CDEA-DQN compared to existing algorithm.