Long-term inspection and maintenance (I&M) planning, a multi-stage stochastic optimization problem, can be efficiently formulated as a partially observable Markov decision process (POMDP). However, within this context, single-agent approaches do not scale well for large multi-component systems since the joint state, action and observation spaces grow exponentially with the number of components. To alleviate this curse of dimensionality, cooperative decentralized approaches, known as decentralized POMDPs, are often adopted and solved using multi-agent deep reinforcement learning (MADRL) algorithms. This paper examines the centralization vs. decentralization performance of MADRL formulations in I&M planning of multi-component systems. Towards this, we set up a comprehensive computational experimental program focused on k-out-of-n system configurations, a common and broadly applicable archetype of deteriorating engineering systems, to highlight the manifestations of MADRL strengths and pathologies when optimizing global returns under varying decentralization relaxations.@en

Long-term inspection and maintenance (I&M) planning, a multi-stage stochastic optimization problem, can be efficiently formulated as a partially observable Markov decision process (POMDP). How- ever, within this context, single-agent approaches do not scale well for large multi-component systems since the joint state, action and observation spaces grow exponentially with the number of components. To alleviate this curse of dimensionality, cooperative decentralized approaches, known as decentralized POMDPs, are often adopted and solved using multi-agent deep reinforcement learning (MADRL) algorithms. This paper examines the centralization vs. decentralization performance of MADRL formulations in I&M planning of multi-component systems. Towards this, we set up a comprehensive computational experimental program focused on k-out-of-n system configurations, a common and broadly applicable archetype of deteriorating engineering systems, to highlight the manifestations of MADRL strengths and pathologies when optimizing global returns under varying decentralization relaxations in such systems.
@en