Deep multi-objective reinforcement learning for utility-based infrastructural maintenance optimization.

Area of Science:

• Artificial Intelligence
• Civil Engineering
• Operations Research

Background:

• Infrastructure maintenance traditionally uses single-objective reinforcement learning (RL), often combining multiple goals like cost and safety into one reward.
• This reward-shaping can oversimplify complex decision-making processes for asset management.

Purpose of the Study:

• Introduce multi-objective deep centralized multi-agent actor-critic (MO-DCMAC) for direct multi-objective optimization in infrastructure maintenance.
• Enable optimization even with nonlinear utility functions, improving upon traditional RL limitations.

Main Methods:

• Developed MO-DCMAC, a novel multi-objective reinforcement learning approach.
• Evaluated MO-DCMAC using threshold and Failure Mode, Effects, and Criticality Analysis (FMECA) utility functions.
• Tested in diverse maintenance environments, including Amsterdam's historical quay walls, comparing against rule-based policies.

Main Results:

• MO-DCMAC effectively optimizes maintenance policies for multiple objectives simultaneously.
• Demonstrated superior performance compared to existing rule-based heuristic policies across various scenarios.
• Validated the method's effectiveness with different utility functions and complex environments.

Conclusions:

• MO-DCMAC offers a significant advancement over single-objective RL for infrastructure maintenance optimization.
• The method provides a more robust and effective approach for balancing competing objectives like cost and safety.
• This research paves the way for more sophisticated and efficient asset management strategies.