Infrastructure · Primitives
ProductivePredictionMachine
A machine simplified by a prediction about its inputs, so that it does the same work with less, for as long as the prediction holds
Definition
Mechanism
Start from an existing machine (a finite automaton, a Turing machine, a program, a process) and a prediction about its inputs, such as 'after an A, only a B ever follows'. Stated negatively, the prediction rules inputs out. Remove the states and instructions that are reachable only if the prediction is false, and the machine gets simpler with the same behaviour. Go further and skip the predicted input altogether: after reading A, jump to the state reached after B without reading B. How much a prediction can be exploited depends on the shape of the machine, i.e. whether parts can be removed without disturbing the rest. This frames prediction as the deformation of an existing shape that already holds information, not as training a blank model toward a reward.
Preconditions
- An existing machine and a prediction that constrains its inputs
Invariants
- Behaviour is unchanged as long as the prediction holds
- The gain comes from states or inputs the prediction makes unnecessary
- How exploitable a prediction is depends on the machine's structure
Risks
Failure modes
The prediction fails and the removed states or skipped inputs were needed
Over-streamlining so that no capacity is left to notice the prediction failing
Never checking skipped inputs, so a broken prediction goes undetected
Relationships
Root pattern: builds on nothing in this library.
Used by
References
Using the room a prediction frees up in a process, either to return visible savings or to run extra functions that the process's outputs do not reveal
Hand back the room a prediction frees up as visible efficiency savings: same function, measurably cheaper, nothing hidden