You've built a policy that scores well. That is not the same as a policy you can trust. Average performance hides exactly the failures a risk manager most needs to find.
Every method so far optimised expected return. But a response strategy that is excellent on average can still behave disastrously in rare, severe circumstances — and in risk management those are the circumstances that end projects and careers. Policy validation is the assurance step: before a policy is trusted with real decisions, you deliberately probe how it behaves in the tail, under model error, and against adversarial conditions. It is the difference between a strategy that looks good and one you'd stake the programme on.
1Finding the rare disasters on purpose
The dangerous scenarios are, by definition, uncommon — so ordinary simulation may barely sample them, and their risk stays hidden behind a comfortable average. Rare-event simulation deliberately over-samples the extremes (then corrects the statistics for having done so) to estimate how often, and how badly, the tail bites. It turns "we've never seen it fail" into an actual probability and severity for the failure you haven't yet witnessed.
2Does it hold when the model is wrong?
A policy is only ever optimised against a model, and every model is wrong somewhere. Robustness analysis asks how the policy performs when reality departs from your assumptions — perturb the transition probabilities, shift the reward, stress the inputs, and see whether performance degrades gracefully or collapses. Related adversarial analysis actively searches for the conditions under which the policy does worst, surfacing brittle failure modes before the world finds them for you. A good policy isn't merely optimal for one model; it's resilient across the plausible ones.
Optimising for the average is not enough. A trustworthy policy is validated where it can hurt you — in the rare, severe tail and under the model errors you know exist — not merely where it performs on a good day.
Part II solved the sequential problem — from the MDP and its exact policies, through approximation and online planning, to policy methods, their stable optimisation, the actor–critic synthesis, and the validation that makes a policy trustworthy. But every one of these methods assumed you knew the model: the transition probabilities and rewards were given. Real projects rarely hand you that. Part III confronts it head-on — deciding well when you don't know the odds, and must learn them while you act.
Never sign off a decision framework on its average performance. Stress-test it explicitly against the rare, severe scenarios that actually threaten your projects, and against the near-certainty that your model is wrong in places. Treat validation as independent assurance, not a formality: a response strategy earns trust by surviving the tail and degrading gracefully under error — not by looking good in the base case everyone already expected.