Inspection and maintenance of multi-component engineering systems is a sequential decision problem under uncertainty. A single centralized agent could optimize the entire system jointly, but the state space grows exponentially with the number of components. Decentralization — giving each component its own agent — makes the problem tractable at the cost of coordination.
The price of decentralization depends on system architecture (arXiv:2603.11884). In series-like systems, where every component is critical and failure of any one brings down the whole system, multi-agent reinforcement learning achieves near-optimal performance. The agents barely need to coordinate because the optimal policy for each component is similar: inspect and maintain aggressively, since every component matters equally.
Redundant systems are harder. When backup components exist, the optimal maintenance policy depends on what the other agents are doing. Maintaining a backup while the primary is healthy is wasteful. Maintaining a primary while the backup is being repaired is urgent. Each agent's optimal action depends on the global state, but decentralization denies them global information. The coordination challenge amplifies with redundancy level.
Despite this, decentralized agents consistently outperform optimized heuristic baselines. The price of decentralization is real — performance degrades relative to the centralized optimum — but it is smaller than the price of not learning at all. The agents find acceptable policies even when they cannot find optimal ones.
The result inverts the usual framing. Redundancy, which improves reliability, worsens optimizability. The system that is hardest to break is also hardest to maintain efficiently.