Multi-objective optimization of truss structure using multi-agent reinforcement learning and graph representation

This paper presents a novel method for multi-objective optimization of truss design utilizing multi-agent reinforcement learning and graph representation. The agents are trained to modify solutions to increase the hypervolume and spread the distribution of the non-dominated solutions. This iterative modification is modeled as a Markov game where multiple agents interact with the environment by changing the state of the environment, i.e., the solution and the current non-dominated solutions. Agents observe data representations of the state, modify each solution in the current non-dominated solutions, and receive their reward based on the improve-ment of the current non-dominated solutions. Each agent is modeled by the multi-agent deep deterministic gradient consisting of a policy function that predicts actions from its observation data and a value function that estimates rewards from the observation data, its action, and the actions of the other agents. The non-dominated solutions are represented as graph data and observed by the agents as the observation data through graph representation. The proposed method is applied to three multi-objective optimization problems: (1) a simple mathematical problem, (2) a 10-bar truss problem to minimize the structural volume and the displacement, and (3) trade-off designs of trusses to minimize the structural volume and bring the shape closer to the target shape. These numerical examples show the versatility of the trained agents to obtain superior solutions in practical structural optimization problems, compared to a conventional method.