Glossary · Machine learning
Oracle
Also called oracle baseline, upper bound, perfect reader
A hand-made baseline that is given information or abilities the learner doesn't have, to show the best result possible.
In machine learning an oracle is a cheat used for measuring, never for the final system. It is given the right answer, or something close to it, and its score shows how well a perfect learner could do under the same rules.
That answers a question learning curves can’t: is the learner failing, or is the task impossible? If even the oracle can’t pass a test, no amount of training will. If the oracle passes easily and the learner doesn’t, the problem is in the learning.
An oracle is only useful if it keeps every limit the learner has except the one it is meant to remove. An oracle that also ignores the learner’s physical limits says nothing about what the learner could reach.
In this project
In the blocking post, the oracle knows which punch is coming. It presses the right block at the opponent’s first visible movement plus the 100 ms vision delay: the earliest a perfect reader could react. The rest is the real body, with its real block speed. Against the hooks and the uppercut it cut the damage sharply, which showed reading was worth learning. Against the jab it did no better than the guard alone: at full speed, nothing reacting to what it sees can block the jab in time.