C.32.LCA:4.9 - Qualify learned contributions by their actual function

Learning is one way to obtain a contribution to the arrangement. Identify which contribution it supplies. A learned state estimate, forecast, action policy, representation or optimization surrogate has different inputs, outputs and failure consequences.

For a predictor, qualify the output, horizon, available observations, future-action assumptions and error relevant to the receiving decision. For a policy, qualify the actions it can select, the states and instructions it can receive, and the constraints of the actual subject. Agreement with recorded demonstrations and a useful closed-loop result are different tests.

Keep training, inference and online adaptation separate. A policy trained slowly can execute rapidly. An online update can change the policy while its input and output formats remain unchanged. Determine whether the relied-on error, timing or constraint result continues to apply to the updated component.

Compare the learned contribution with an adequate programmed or conventional controller at the same receiving use. Include data acquisition, training or calibration, inference resources, monitoring and recovery in the whole cost. Prefer the learned option when its supported capability changes the result enough to justify those burdens, not because the architecture contains a neural network.

If a separate filter or supervisor modifies a learned command, qualify the resulting combination. The modified command can preserve one constraint while changing a tracking or progress result. A human fallback also needs detection, sufficient time, usable information and actual authority.

The output of this operation is a qualified component use or a named missing result. It is not a general method for training every controller. When the application needs a bound that the learned supplier has not established, compare a narrower operating region, an applicable protective construction, a different supplier or a weaker warranted claim. Do not silently replace the missing bound with average benchmark performance.