C.40:4.12 - Develop and use materially different continuations

Use this branch when improving one kind of result is eliminating other worthwhile possibilities, or when retained variants still provide no usable next action. The result is material that supports different examined uses or further development, together with the operation that makes those differences usable. Start with a few explained variants when that is enough. Maintaining a large collection, measuring novelty or building an ensemble adds work and needs its own reason.

Choose which differences to preserve. Begin with the use that requires alternatives: handling different situations, exploring ways whose eventual use is unsettled, preserving an intermediate construction, or combining complementary behavior. Recover the conditions and observations that make a difference consequential. Two implementations can produce the same behavior now yet permit different future changes. Conversely, very different encodings can produce the same behavior. Describe the property that matters at the point where the next operation uses it; inspect actual applications when the property depends on their outcome. C.17 supplies a qualified characterization, and :4.11 supplies the connection between representation and available change.

For example, an image’s encoded byte length need not distinguish the visual forms the designer wants to retain. A controller’s parameter distance need not distinguish its responses in the situations that matter. A behavior observed on one input supplies that case; it cannot characterize unseen situations by itself. Use trajectories, responses to selected inputs or another obtainable observation when the receiving difference requires them. Retain a consequential rare failure beside an aggregate. If the chosen observations cannot expose the difference, change that examination before using its score to protect diversity.

Then choose how much variety to keep and how to compare like with like. A few named kinds can suffice. A grid over descriptors is useful when its cells distinguish relevant behavior and its resolution is affordable. A distance-based neighborhood can avoid a rigid grid but needs a distance and scale that express the intended similarity. Adding coordinates or finer cells can multiply the material and examination needed without changing a useful choice. Empty regions remain unproduced regions unless a separate argument establishes impossibility.

Compare keeping separate ways with one way that takes the relevant condition as an input and produces an adequate response. A parameterized construction, a goal-conditioned controller or direct enumeration may answer the same need with less retained material. Include its training, planning, observation and use costs. The smaller description is a saving only if that complete arrangement supplies the required differences.

Connect production, local comparison and renewed use. When local competition is worthwhile, construct the following loop with the available professional operations:

1. Obtain initial feasible candidates and retain their actual editable or executable material. Include a sufficient incumbent when one exists. State the behavior distinctions, comparison conditions and quality criterion used in this pass.
2. Select retained material and perform a permitted variation or recombination. Selection can draw from another behavior group: useful changes need not remain in their parent’s group. If no existing material can support the attempt, obtain another feasible starting candidate.
3. Apply or inspect the new candidate to establish its behavior and quality on the declared basis. Assign its group from that result. A requested descriptor is a target for generation, not an observed descriptor.
4. For an empty relevant group, retain a usable candidate. For an occupied group, compare with the incumbent there and replace it only when the resulting retained set better serves the declared use. Protect essential conditions and material needed for another continuation. In the simple one-winner arrangement, retain the better local candidate and resolve ties without pretending that equal scores establish interchangeability.
5. Use the retained material to choose the next variation, actual application, reexamination or stop. Keep enough construction, inputs and conditions for that chosen operation to be performed.

This is the useful connection in a MAP-Elites-style construction. A higher global score no longer expels every other kind of behavior. The group’s winner remains the best retained candidate under that comparison, with its stated limits. The original MAP-Elites algorithm, §3, supplies that particular grid arrangement. The same receiving question can warrant neighborhoods, several retained candidates or a direct small set instead.

Selecting parents and retaining results answer different questions. Uniform parent selection gives each retained candidate an opportunity but dilutes that opportunity as the collection grows. Bias toward parents whose recent variations improved the collection can focus work, but that success depends on the current collection and attempted operators. Preserve an affordable way to examine an uncertain region when its possible continuation matters. Retention alone supplies no opportunity if a parent is never selected or all its descendants are filtered out.

Preserve exploration without confusing it with present quality. A low-scoring intermediate may support a useful later change. Protect its actual material and a bounded opportunity to investigate that continuation. When the eventual use is unsettled, the reason can be an intelligible exploratory question about a region or kind of behavior; it need not assert the value of a particular unknown descendant. Stop when the question no longer warrants its burden.

Novelty-guided search instead favors behaviors different from a declared reference set, which can include the current population and selected history. Specify what observation, distance and reference update make the difference meaningful. Recompute a novelty judgement when its reference changes. A history used to guide exploration can need different retention from the compact collection returned for use. Local quality comparison may be combined with novelty, but maximizing novelty alone supplies no quality guarantee.

Crowding control and separated subpopulations are other ways to limit premature concentration. Their effect depends on the similarity account, opportunities for variation and exchange between groups. A rule that repeatedly replaces a boundary candidate with a slightly better neighbor nearer the center can erode the very range being preserved. Examine the resulting collection, not only each replacement’s score. If a promising path is lost, distinguish an unavailable change, an early filter, parent neglect and a retention rule that removes its intermediate. Repair that connection rather than merely increasing the number of attempts. Sections 4.7 and 4.11 develop those returns.

When a population has become too similar to provide useful further variations, restarting around a retained useful result can reopen the search. Preserve that result as a fixed base and construct a fresh population of permitted changes relative to it. For numerical material, obtain each candidate as base plus change, then examine the resulting candidate; a small change value is not itself a good result. Include the unchanged base for comparison and choose the range of changes for the question, with a way to explore beyond small adjustments when useful. After a justified improvement, save the resulting material as the new base before another restart. For other material, supply the actual edit-application operation instead of assuming numerical addition. This delta-coding construction can restore variation; it cannot repair an unavailable operation, uninformative evaluation or an exhausted useful question. Compare it with an independent fresh start, a different representation and stopping.

Use repeated contraction and regrowth only for a justified exploration question. Another arrangement changes which lines survive by periodically reducing an active experimental population and growing it again. It is worth comparing when existing search concentrates its descendants into too few useful kinds. A restart around a fixed incumbent, continued search without contraction and the direct continuation trial in :4.11 remain alternatives.

Construct the full cycle before attributing a benefit to the reduction. Define the behavior groups, active population, permitted reproduction or variation, population bound and affordable interval between events. One explicit survivor rule chooses a small stated number of occupied groups uniformly without replacement and keeps all their members. This differs from choosing individuals uniformly: group size no longer sets the group’s chance of selection. Other rules, such as keeping one representative from each selected group or always retaining a champion, change the opportunities and must be specified as such.

After contraction, produce new candidates from the survivors, examine them and admit them under the stated quality and diversity conditions. Allow the population to refill: a replacement rule that cannot increase its size needs a temporary refill rule, with its admissibility conditions retained. Continue variation for the declared interval, then apply the next survival event to the groups now actually occupied. If the survivors cannot reproduce or admissible variation cannot refill the population within the allowance, return that failure rather than treating planned regrowth as achieved. The interval must leave an opportunity to generate a spread before the next event; no universal event frequency follows from the construction.

The connection to future variability is conditional. Suppose four groups are occupied and the rule keeps all members of one uniformly selected group. A line with descendants in three groups has survival probability 3/4 at that event; a line confined to one has probability 1/4. Wider occupation changes survival under this rule. For repeated events to favor a continuing ability to spread, the survivors must carry material that can produce another spread during regrowth. One lucky placement, one deletion, or better current fitness does not establish that ability. The argument also changes when group definitions, survival rules or inherited changes change.

Compare the complete cycle by the useful results and total effort obtained, including lost lines and rebuilding work. Where the work permits it, retain useful delivered answers outside the active experimental population; deleting an experimental parent need not destroy an already usable result. Reintroducing saved material or protecting elites changes the cycle and its comparison. When a learned descriptor changes after contraction, recharacterize retained candidates through the following construction before using the new groups. Contraction can change the data used to learn that descriptor, but a changed map alone is neither new behavior nor evidence of improved future variability. Stop or change the arrangement when destruction and regrowth cost more than the continuations they obtain.

Reexamine the collection when its basis changes. Chance can affect both quality and which behavior is observed. One lucky trial can then create a false local winner or apparent coverage. Choose the further observations that could change the retention or use decision and reserve enough effort for them. Keep evidence for the actual candidate and conditions. Expected performance and repeatability are different receiving requirements; a mean score cannot settle both.

When reexamination changes a candidate’s descriptor, withdraw its old current membership, update its supported characterization and place it under the current retention rule. Reconcile any collision with another candidate. Retain the old observation as history when useful, rather than counting it as an additional currently supported behavior. Temporary alternatives within a group can preserve a fallback when an apparent winner fails. Their carrying and rechecking costs count alongside new trials. Repeated observations improve the stated estimate; they do not remove a shared measurement defect or establish behavior in an unexamined regime.

A revised descriptor or learned representation creates a related return. Keep the material and observations needed to characterize retained candidates on the new basis; recompute their descriptors and rebuild the affected grouping and comparison. Merely changing the threshold for future additions leaves old members under old conditions. If the required observation was never retained and cannot be recovered, mark that part unresolved or examine the candidate again. A rearranged map alone is no new behavior or coverage achievement. Keep a means to explore beyond the regularities in the material used to learn the descriptor or generator.

Make the retained difference available to the next operation. Select the actual receiving branch:

• For a choice among alternatives, return the relevant candidates, their differences and supported conditions; make the receiving choice through its own Method. C.18 supplies an archive or front claim when required. C.19 supplies policy for continuing, retaining, narrowing or retiring live lines.
• For another generation, supply the retained construction and the changing operation. Preserve relationships the next change needs. If two binary choices must sum to one, independently sampling them from the previously useful pairs (0,1) and (1,0) can produce inadmissible (0,0) or (1,1). Sampling a whole pair preserves that known relation; producing a useful new combination requires a further relation-preserving operation. A learned generator similarly needs its actual fitting and generation Method through :4.11, followed by examination of its proposals. Using more historical material can reduce one loss while increasing fitting, storage and stale-condition costs.
• For one combined way, construct how the components receive inputs and how their outputs produce an action. A vote, numerical average, selection rule and combination of action parts are different operations. Check the resulting behavior: averaging two actions that pass on opposite sides of an obstacle can send the combined action into it. Train or develop a selector through :4.7 when its rule is not already obtainable, then compare the complete arrangement with a sufficient single way under common conditions.
• For learning from retained behavior, supply the relevant situation, actual action and outcome, the rule for selecting useful examples, and an available learning operation. A weak overall performer can supply an excellent local example. The recipient still needs the observations, action meanings and capability to learn from it. Apply the operation and examine the recipient’s resulting behavior. Sections 4.8 and 4.10 preserve the additional conditions for expert and human contributions; a transmitted example does not itself transfer a capability.

When learning is part of a candidate’s examination, choose what goes into subsequent candidates. One arrangement uses performance after learning to select the original inherited material: descendants constructed from it still need their own learning operation. Another writes compatible acquired changes into the material that descendants inherit, such as trained weights in an editable network encoding. Preserve the original, learned state and inherited material distinctly, and perform the chosen copying or reconstruction. A learned state that cannot be encoded there requires another construction. Section 4.7 supplies the fresh-start and continued-state comparisons; include the learning needed by descendants rather than crediting them with the parent’s acquired performance before they can produce it.

For a combined stateful way, follow the history actually performed. A component that was idle may have stale state; one updated on its own imagined actions may have the wrong history. Supply the real history or a supported state transfer, redesign the component to use available state, or retain the resulting gap. Reinitialization is sufficient only when it preserves the intended use. Test the switching and combined trajectory, not merely each component used alone. Availability of the selector’s inputs at the moment of action is part of this construction.

If the components themselves need further development, use their participation in complete ways to guide it. Assemble compatible combinations, perform and judge each whole, and associate its outcome with the participating versions, partners and connection. Compare a proposed component change with the incumbent in common relevant combinations; try other partners when a complementarity question warrants it. A component that performs poorly alone can be useful in such a combination. Conversely, a better isolated component can make the whole worse. An average over its sampled participations is evidence for those combinations, not an intrinsic value or an isolated causal contribution. Reexamine the affected wholes when partners or the selector change. This feedback can favor complementary parts, but does not guarantee that diversity will maintain itself.

Return the material and supported continuation, including a sufficient direct answer when the larger arrangement adds no worthwhile result. Count generation, repeated examination, learning or selector development, storage, recovery and recurring use. Retiring active development can leave inexpensive useful material retained; an obsolete or unaffordable continuation can justify retiring that retention too. The difference made in practice is that a developer can reproduce how useful alternatives survive, how they are used again and which connection a changed condition reopens.