CMP.Reference:1.4 - Descriptions can be computed with, and ways can be obtained
A computational description can be represented as data. An interpreter executes its rules; a translator produces another description; an analyzer obtains a claim about possible executions. These are different results. CMP.12 constructs interpretation and meaning-preserving translation, while CMP.13 constructs a sound approximation for an analysis question. Binding, state, control and the chosen observations determine their connection.
This enables more than repeated execution. A search can explore descriptions of candidate algorithms, and a learner can obtain a rule from examples or feedback. The outer computation needs an effective candidate representation, ways to propose or change candidates, and a criterion it can evaluate. The resulting rule has another use with its own inputs and consequences. CMP.4 supplies justified search exclusions; CMP.6 develops updates from local information; CMP.7 connects the learner, its selected rule and further use. A good update is one contribution to a learner, not a guarantee about every later prediction or action. The Suite Reference’s §3.10 develops this connection for learned controllers and changing computational organizations.
Universal execution, approximation, learning and practical accessibility answer different questions. A universal interpreter can simulate the computations expressible in its admitted model; that claim does not supply a solver for every stated problem or a favorable runtime. An approximation theorem can establish that a representation of sufficient quality exists while leaving its size and obtaining procedure unsettled. A learner’s useful generalization depends on restrictions and evidence connecting observed and future cases. A physically available arrangement adds its own preparation, accuracy, energy, storage and communication conditions. Do not use any one of these claims as a substitute for the others.
General search and learning can reduce how much behavior must be specified for each case. They still embody choices of representation, feedback, interaction and resources. More computation can improve a result only through a way that can use it; it cannot recover information absent from the available observations or validate an unrelated objective. A failed attempt can therefore call for another algorithm, another representation, a better source of evidence or a changed useful question. It need not call for a larger run of the same method.