C.40:4.7 - Develop a way by what its use produces

Use this branch when the material being varied is a way of obtaining a result: a learning rule, a constructive heuristic, a training loss, a procedure with adjustable settings, or executable code. The useful difference appears only after that way has been applied. Choosing a good finished schedule and discovering a rule that makes good schedules on further inputs are different searches. A sufficient known way, direct derivation or small complete comparison can avoid the larger search.

Construct the two connected operations. Give the candidate a usable representation and a permitted changing operation; section 4.11 develops that connection when it is missing or obstructs a needed variation. Parameters permit tuning within a form; changing a formula, operation or connection can enlarge the forms reachable. Include a known usable incumbent. Start with variations whose application can actually be performed and examined; a larger representation is useful only if it exposes a consequential alternative at an affordable search cost. C.39 develops an unavailable obtaining operation, and C.38 compares serious complete-enough alternatives.

For each candidate, perform the obtaining operation on stated inputs, then judge the result it produced. In a learning application, this includes initialization, training and subsequent use of the trained model. In a scheduling application, it includes constructing and examining a schedule. Recover the candidate way, its starting state, the resulting artifact or state, and the outcome separately. The outer development changes the way; the inner application obtains the result used to compare ways. These are roles in this particular construction, not a universal hierarchy of Methods.

Specify which inputs the candidate may use and which observations its assessment will use. A learning rule can see training examples while its performance is assessed on other examples. Developing a reusable way across situations also needs different situations: an answer hard-coded for one task can pass that task without learning or constructing anything useful for the next one. Protect important failure classes beside any aggregate. Keep cases repeatedly used to select changes distinct from a later examination supporting a stronger receiving-use claim. Repeatedly consulting a reserve makes it selection material.

flowchart LR
  A[Candidate way and starting conditions] --> B[Apply the way to task inputs]
  B --> C[Produced artifact or changed state]
  C --> D[Judge the receiving result]
  D --> E[Choose a variation, retain, or stop]
  E -->|continue with changed material| A
  F[Cheap forecast or reduced trial] -->|chooses what to examine| B

The arrows show result use and a return for development. The forecast selects work; the produced result supports the corresponding outcome claim. When the use is prediction alone, C.29 supplies that different bounded claim.

Choose what is allowed to change together. A useful form may require its own tuned parameters. Compare a form with parameters fitted by the allowed inner operation rather than condemn it under settings suited only to its rival. Conversely, adding adjustable parameters is not automatically an improvement. Compare fixed forms, tuned forms and changed forms where those alternatives can settle the question. State the allowed tuning work for each; count it in the development burden. If the changed form needs another initializer, data preparation or support, compare that complete arrangement and preserve the dependency in the returned way.

Changing a training loss changes what guides updates. It does not by itself change what the receiving use values. An outer comparison can favor a loss whose numerical value is larger or whose early learning is slower because the resulting model performs better in the intended use. If the evaluation itself is defective, return to A.19.ECS and E.23; optimizing the candidate against that defect is no repair. A genuine change of the receiving requirement instead reopens the affected question through C.40.CD.

Make search affordable by choosing an explicit approximation. First obtain enough actual applications to make the comparison intelligible. Then choose among the following arrangements according to what is costly and what evidence the decision needs. They can be combined when their combined losses remain acceptable; they are not compulsory stages.

For prediction-assisted search, choose features available at the time the decision must be made. Relate them to actual completed outcomes from the relevant regime; a feature measured only after completion cannot save that completion. Fit the prediction rule, use it to propose the next costly trial, perform that trial, and update the rule with the obtained outcome. An early trace can support a forecast of later performance but can miss a late improvement. Keep a way to investigate such misses when they matter: complete a discriminating delayed case, reserve some effort for uncertain regions, or broaden the calibration data. Stop relying on the forecast for a changed regime until its relevant relation is supported.

An early stopping rule similarly rejects further work, not the possibility of eventual success. If it is calibrated only on fast learners, a repeatedly discarded slow family can remain invisible. Recheck that omission against the receiving question. A cheap filter for malformed candidates is different: it can reject a proved type error or inadmissible operation without predicting final quality. Reusing an earlier result for an equivalent candidate requires the relevant equivalence; agreement on a few sampled outputs provides only that sample’s evidence.

For asynchronous evaluation, retain the identity and starting conditions of every pending candidate. Queue enough independent work to use the available workers. When a selected batch of results returns, compare those results under their applicable conditions, select material for variation, and submit replacements without discarding the still-running candidates. A small return batch permits prompt adaptation; waiting for a larger batch can reduce how strongly a few fast lineages determine the next proposals. Retain each result’s task, starting state, work allowance and candidate version. Check whether older parent material or faster return changes what is selected, and refresh the material used for variation when that is the cause of lost progress. Compare resulting quality and total work as well as elapsed time before retaining the asynchronous arrangement.

Choose a fresh-start comparison or continued development. With independent applications, each candidate receives a defined starting condition and work allowance. This supports a claim about the candidate’s way under those conditions. With continued development, a successful state can be copied or retained, its settings changed, and work resumed. This can produce a better final artifact cheaply even when the last settings would be poor from the original start. Return the trajectory or reproducible continuation, including the state and earlier operations it depends on. A separate fresh-start trial is needed only for the stronger claim that the final settings define a reusable way from that start.

For example, in a stipulated update toward a known target 10, the rule is x ← x + a(10 − x). Two updates with a=0.5 from x=0 produce 7.5. Two updates with a=0.1 from an inherited x=8 produce 8.38, but the same rule from x=0 produces only 1.9. The continuation ending at 8.38 has an error of 1.62, smaller than 2.5 for the other run; this supports that continuation. It does not show that a=0.1 is the better two-step rule from zero. Directly setting x to the already known target would be cheaper in this toy problem, so the arithmetic illustrates the comparison distinction, not a need for evolutionary search.

State reuse needs a legitimate realization. Software may permit copying model weights and compatible optimizer state. A changed structure may make that state unusable. A person’s acquired skill, fatigue or experience cannot be cloned by copying instructions. For development involving people, use actual preparation and learning conditions through HCD and actual trials through ME.11. If a comparison requires identical acquired states, keep that unavailable condition explicit; copying instructions supplies no such state.

Develop persistence and recovery from states the process reaches. A repeated rule can produce a wanted result and then destroy it, or work from its original start but fail after a disturbance. Use this construction when the receiving demand includes continued useful operation or recovery. Distinguish what must be obtained from the original start, what must remain true during use, and which disturbances must be recoverable within the available time and resources. Continued functioning can involve changing configurations; a fixed picture need not be the target.

Begin with a sufficient known rule or complete small comparison when available. Otherwise, one arrangement is to run candidate rules for longer and judge the relevant result at several times, using that feedback to change the rule. Checking only the first attainment of the target leaves later destruction unexamined. Checking only one later endpoint can miss an intervening failure. In gradient-based training, retaining a long computation for a backward pass can be expensive; other forms of development can have different costs.

When restarting from retained states makes the work practical, construct a small pool of continuation states. Each entry contains what the process needs to resume, including necessary hidden state and relevant conditions. It must be copyable or reproducible by available means; a picture or score alone may not provide it. Use actual long trials or obtain the missing continuation support when such starts cannot be realized. The rule being developed, one state it produced and the useful whole obtained through execution remain distinct.

Initialize the pool with the original permitted starts. For a development trial, take a manageable selection from the pool and include an original-start case. Run the current rule for an agreed interval, inspect its produced state and receiving result, and use that feedback to update the rule through the available learning operation in CMP.7 or to compare executable variations. The feedback can concern a whole later configuration or function; it need not supply the correct local action at every intermediate step. Return suitable output states to the pool as starts for later trials. Keep enough original-start trials to expose loss of the ability to begin, as well as reached, incomplete or failing states whose continuation matters. Merely keeping a seed in storage gives it no influence on development.

The returned states make the next trial begin where an earlier application ended. A rule must now maintain or improve an already formed result as well as create it. Limit the pool and select its entries by the needed state range and affordable work; keeping only easy successful states can hide a recurring failure. Record which rule and conditions produced an entry. If a revision changes what its stored state means or requires, reconstruct compatible starts or examine a supported conversion before reusing it. Restarting from a saved state avoids storing its entire preceding computation; it does not assert that the resulting learning update is equivalent to training through that whole computation.

For recovery, apply a relevant permitted disturbance to some suitable states before continuing their trials. Keep undisturbed and original-start cases where those abilities remain required. Perform the continuation, judge recovery within its allowance, and keep running long enough to examine whether the recovered result functions and persists. Use consequential failures to change the rule or the conditions it needs, then repeat from the affected starts. Match the training feedback to the receiving result: resemblance to a desired body may support a shape claim while its subsequent movement remains poor. Compare actual functioning when that is what the work needs. States and disturbances repeatedly used to choose changes are development material; examine further relevant conditions before making a broader claim.

Damage can also remove the information that identifies the intended result, or remove the means to restore it. Obtain a known target, retained distinguishing information or qualified assistance when different required restorations are compatible with the same remaining state. Obtain the required actuation, material and energy for a physical repair. If those contributions are unavailable, return that limit; additional training cannot perform an absent operation. Stop with a sufficient directly chosen rule, a supported growth-and-continuation arrangement, or the exact missing contribution. Finite trials support their examined conditions; persistence for an arbitrary duration needs an additional ground, such as a preserved invariant.

Learn from the histories the developing way causes. Use this construction when a learned action rule or fitted component helps determine the inputs it will encounter next. A small initial error can take it outside the histories demonstrated by a competent source. Keep a fixed dataset when it already supports the intended behavior; use a direct rule when that obtains the required response more cheaply.

Run the candidate inside the permitted interaction, retaining the observations and actions it actually causes. At a consequential reached history, obtain a teaching target from a suitable expert, source model or other qualified feedback operation. The target can be a corrected action, a distribution over continuations or a component output that supports the acting way. Pair it with the information the learner will actually receive. An observation or outcome following the old action belongs to that action; obtain the response to a corrected continuation through a new permitted application or supported model.

Fit the selected learner through CMP.7, then run the changed candidate in the interaction again. Use the new histories to obtain further needed targets, retaining relevant earlier examples and protected behavior. A changed learner changes what will be encountered; repeating the original expert demonstration alone can miss the new error. Stop when the required behavior is supported at its receiving conditions or the remaining learning no longer warrants its collection and use costs. More collected histories are useful only through the learning and later behavior they permit.

Training may use information unavailable during action, such as a simulator’s hidden state, to obtain targets. Keep that information out of the learner’s operating input unless it will be available there. Test whether the available history can support the required response. If indistinguishable histories require different actions, return to the observation, retained state, supported assistance or narrower use; inconsistent targets cannot repair the missing distinction. MMP.8.SD supplies that information-and-timing construction.

Check whether the teacher can supply useful targets at the learner’s reached histories. If it needs adaptation, obtain several teacher continuations from a learner-produced prefix. Judge their completed outcomes with an available check, use that feedback to train the teacher on its continuations, and return the updated teacher to the learner’s next training. Repeat as the learner changes when benefit warrants the cost. This requires access to change the teacher and useful outcome feedback; otherwise retain another supplier or an unsupported continuation. Improved completed outcomes do not establish correctness of every local instruction. Include data collection, target preparation, fitting and repeated interaction in the complete development comparison. A source-model trial supports that model’s setting; physical or other receiving use still needs its own permitted examination.

Return a usable discovery and its supported claim. Keep the executable construction or sufficiently developed description, necessary initialization and support, settings or adaptation rule, outcome comparison, and consequential failures. If the sought result is the final trained artifact, return that artifact rather than claim a new reusable learning way. If a general way is sought, try it on the relevant different inputs and examine accidental dependencies on the search setup. Separate changes that preserve the computation from simplifications or decouplings whose effect needs a new comparison. Removing an apparently redundant instruction can expose a useful compact method; a component that merely looks strange may carry the improvement. Test the changed construction before replacing it.

Compare development cost, recurring use cost and time to a useful result separately. Count candidate construction, partial and full trials, tuning, repeats, prediction-model construction and updating, retained state, and the receiving examination where applicable. Fewer full trials can coexist with expensive feature collection; less waiting can coexist with more total work. Stop with a sufficient incumbent or retained conditional alternative when further improvement does not warrant that burden. Changes to task inputs, training horizon, allowable support or receiving criterion reopen only the comparisons they can invalidate.