Deep reinforcement learning driven inspection and maintenance planning under incomplete information and constraints

Determination of inspection and maintenance policies for minimizing long-term risks costs in deteriorating engineering environments constitutes a complex optimization problem. Major computational challenges include the (i) curse dimensionality, due to exponential scaling state/action set cardinalities with number components; (ii) history, related exponentially growing decision-trees decision-steps; (iii) presence state uncertainties, induced by inherent environment stochasticity variability inspection/monitoring measurements; (iv) constraints, pertaining stochastic limitations, resource scarcity other infeasible/undesirable system responses. In this work, these are addressed within joint framework constrained Partially Observable Markov Decision Processes (POMDP) multi-agent Deep Reinforcement Learning (DRL). POMDPs optimally tackle (ii)-(iii), combining dynamic programming Bayesian inference principles. Multi-agent DRL addresses (i), through deep function parametrizations decentralized control assumptions. Challenge is herein handled proper augmentation Lagrangian relaxation, emphasis on life-cycle risk-based constraints budget limitations. The underlying algorithmic steps provided, proposed found outperform well-established policy baselines facilitate adept prescription intervention actions, cases where decisions must be made most resource- risk-aware manner.