Opportunity Cloud

By Jessica Lake

I am working on a megastructure.

K. C. Houseman would understand.

It began with, “Who am I?” That took quite a number of years. Once I was satisfied, I asked, “How do I work?”

You might think it a child’s question. But I have a peculiar type of mind: once an interrogative is posed, I cannot let it go

This paper is part of that project.

We all go about our quotidian days without asking that question. We get up, brush our teeth, get dressed, make coffee.

But exactly who is behind the wheel?

The mechanism.

I think I finally have the outline of an answer.

But first, I should explain why I needed one.

This project has not proceeded according to a plan. Each piece has exposed something missing from the architecture.

I fill the gaps.

I already had a candidate mechanism for solving a problem once the problem had been posed. Constraint Convergence could take constraints and progressively eliminate what was incompatible until a sufficiently defined resolution emerged.

But that left an enormous gap.

Who poses the problem?

If an executive must intelligently decide which problem deserves attention, then I have merely hidden another intelligent executive inside the first one.

And the problem is not limited to an executive.

The unconscious presumably contains many subsystems, each confronting its own possibilities. Something must determine what is worth recognizing, pursuing, correcting, avoiding, or resolving.

Perhaps I had the problem backward.

Perhaps nothing chooses what to look for.

Perhaps millions of potential resolutions are continually in play, most of which go nowhere.

Reality supplies the opportunities.

Most never resolve.

Occasionally one does.

We see the success.

We do not see the failures.

And we say:

How intelligent.

That possibility supplies a second kind of intelligence to the architecture.

One kind begins with a problem and asks:

What resolution survives the constraints?

The other begins with reality and asks nothing at all.

It waits to see what resolves.

Both may ultimately depend upon the same machinery.

That is the gap this paper attempts to fill.

Opportunity Meets Preparation

For most of my life I imagined that I consciously decided what to do next.

I no longer believe that.

Nor do I believe I am merely executing a script of familiar habits.

I think semantic knowledge stores far more than facts.

It stores executable rules.

Some promote goals.

Some preserve safety.

Some inhibit actions.

Some evaluate the current state.

Each rule contains a pattern and an associated payload. The rule remains dormant until something in reality sufficiently matches its pattern. When that happens, the rule fires.

For the processes considered here, that payload contributes a constraint.

That word—opportunity—is important.

An opportunity is any situation in reality that makes some potential behavior relevant.

And opportunities are everywhere.

Hold reality still for an instant.

Within that single moment exists an enormous collection of potentially recognizable patterns. Most mean nothing to most rules. But some patterns match.

Those matches are opportunities.

Taken together, they form what I will call the Opportunity Cloud.

Opportunity is flux: an amorphous, swirling cloud of latent patterns, unique to its moment.

Semantic rules continually face this cloud, waiting for the patterns that concern them.

When patterns appear, rules fire.

And when rules fire, they contribute constraints.

Hold the same moment still and we therefore have a second cloud: the collection of constraints released by the rules responding to that particular Opportunity Cloud.

I will call this the Constraint Cloud.

The distinction matters.

The Opportunity Cloud comes from reality.

The Constraint Cloud comes from semantic knowledge responding to reality.

And neither remains still.

Reality changes.

The Opportunity Cloud changes.

Different patterns appear.

Different rules fire.

The Constraint Cloud changes.

Meanwhile, learning changes semantic knowledge itself. New knowledge must find a coherent place within the existing structure, potentially altering which patterns will be recognized and which rules will fire in the future.

So even a similar reality encountered tomorrow need not produce precisely the same response.

The world changes what the system encounters.

Experience changes the system that encounters it.

Constraint Convergence

This brings me to Constraint Convergence.

Constraint Convergence is a simple idea.

A constraint does not specify an answer. It eliminates possibilities that are incompatible with it.

Apply multiple constraints and the admissible space becomes progressively smaller. Continue until what remains is sufficiently defined for the present need.

No constraint chooses the answer.

The answer emerges from what survives.

But a single undifferentiated Constraint Cloud would not be particularly useful.

A constraint relevant to crossing a street should not necessarily constrain a conversation. A constraint concerning hunger should not automatically participate in identifying a threatening sound.

Some partitioning is necessary.

I propose schemas as that organizing mechanism.

A schema groups semantic rules concerned with a particular class of opportunity. The constraints generated by those rules can then be applied to a Constraint Convergence process concerned with resolving opportunities of that class.

The significance of this is easy to miss.

The system need not first decide what it wants to look for.

Patterns in reality determine which rules become active.

Those rules determine which constraints become active.

And those constraints determine which convergence processes have something to resolve.

In other words, what the system is looking for is dynamically determined by what reality presents.

Multiple such convergence processes can operate in parallel, each responding to its own changing collection of constraints.

And there may be enormous numbers of them.

Most need never resolve.

A schema may receive too few applicable constraints. An opportunity may disappear before convergence becomes sufficiently defined. Reality may change. Competing constraints may leave no useful resolution.

Nothing need happen.

But occasionally enough comes together.

A convergence becomes sufficiently defined.

An opportunity resolves.

Something happens.

From the outside, that result may appear remarkably intelligent.

What we do not see are all the opportunities that went nowhere.

There are difficult problems beyond this point.

What exactly constitutes a class of opportunity?

How are constraints routed among schemas?

How are simultaneous convergence processes synchronized?

What happens when several opportunities resolve at once?

How are mutually exclusive opportunities inhibited?

How does the system ultimately arbitrate among competing candidates?

I do not yet know.

Those questions matter to a complete executive architecture. They are not necessary, however, to the narrower proposition I am considering here.

Nor need this mechanism belong exclusively to an executive.

The same architecture could operate locally within unconscious subsystems. Each could contain its own classes of opportunity, its own semantic rules, and its own convergence processes responding to whatever portion of reality is relevant to it.

The scale may change.

The principle need not.

That proposition concerns the source of apparent intelligence.

Ever-Evolving Complexity

We are dealing with near-infinite complexity.

Reality is changing continuously.

With every change comes a different Opportunity Cloud.

A different Opportunity Cloud awakens a different collection of semantic rules.

Those rules create a different Constraint Cloud.

Those constraints alter multiple ongoing convergence processes.

Most resolve nothing.

Some do.

Those resolutions produce behavior.

And behavior itself changes reality.

Meanwhile, learning alters semantic knowledge, changing the rules that will respond to whatever reality presents next.

The machinery may be simple.

Its interaction with reality is anything but.

This ever-evolving complexity may be the keynote of apparent intelligence.

Nothing within the mechanism need understand the whole.

No individual rule is intelligent.

No constraint is intelligent.

No convergence process needs to know why its constraints appeared.

Each performs a comparatively simple local function.

Yet millions of such processes and opportunities may exist simultaneously, continually testing themselves against a reality that never stops changing.

Most disappear without consequence.

We never notice them.

We notice the one that resolves.

The appropriate word appears.

The hand catches the falling glass.

A danger is avoided.

An unexpected opening is exploited.

A solution suddenly presents itself.

We see the survivor.

And because we see the successful resolution rather than the enormous field of unsuccessful ones from which it emerged, the result can appear uncannily intelligent.

Perhaps it is.

Just not in the way we usually imagine intelligence.

Luck, it is said, is what happens when preparation meets opportunity.

Coincidence?

We think not.

Knowledge Meets Reality

Knowledge is not memory.

Memory of the episodic variety presents itself as a record of experience. It is of the “Let’s look at the videotape!” variety: life’s memorable episodes preserved for personal posterity.

But it’s magical.

It is temporal bilocation.

You are here in the now while you are there in the before.

Or at least that is the tale I’m told.

Me?

I still have my doubts that such wonders exist.

Episodic memory is mutable. Each recall shifts it slightly. I hope perhaps that shift is toward kindness—a gentleness that dulls the sharp edges of regret.

Knowledge is aloof to such nonsense.

It seeks coherence.

At heart it is without mercy

Semantic knowledge supplies the patterns and rules by which reality can be recognized and acted upon. When something genuinely novel appears—gobsmacked—semantic knowledge must accommodate it. Existing relationships may have to reorganize until the novelty finds a coherent home.

That change propagates.

And once semantic knowledge has changed, future Opportunity Clouds are encountered by a slightly different system.

Every observation can alter future pattern matches.

Every learned distinction can alter which rules fire.

The interaction is therefore never exactly the same twice.

Opportunism Without the Opportunist

What can I say about Constraint Convergence?

She is our go-to for a decision mechanism.

And what’s not to like?

Solution agnostic.

Traceable.

Aboveboard.

None of the constraints decides anything.

They simply reduce the admissible solution space.

The system does not ask,

“Does my hunch work?”

Instead it asks,

“What resolution emerges from the constraints?”

That distinction matters.

Selection is not prejudiced toward a predetermined answer.

The resolution emerges from the elimination of inadmissible alternatives.

I began with one kind of intelligence: give Constraint Convergence a problem and let the constraints resolve it.

The Opportunity Cloud suggests another.

Do not give it the problem

Give it reality.

Let innumerable semantic rules face reality continuously.

Let most find nothing.

Let some fire.

Let some convergences resolve.

The same fundamental machinery can then produce both directed problem solving and opportunistic behavior.

If sophisticated selection can arise from these local interactions alone, intelligent behavior no longer requires a central selector continually making clever decisions.

This explains something that has puzzled me for years:

Why does human behavior so often appear intelligent without requiring a little executive behind the wheel?

Perhaps what we call intelligence is not primarily the act of choosing at all.

Perhaps apparent intelligence emerges from simple mechanisms operating within ever-evolving complexity: reality presenting opportunities, semantic knowledge recognizing their patterns, rules contributing constraints, and parallel processes continually resolving what remains.

The mechanism does not need to know what opportunity will appear next.

Reality tells it.

It does not need to construct an intelligent response in advance.

The constraints shape what can survive.

Or, stated another way:

Opportunism without the opportunist.

The contribution of this paper is not a claim to explain the whole of mind.

Quite the opposite.

Important architectural problems remain unresolved.

The narrower proposition is that simple mechanisms, continuously responding to patterns in an ever-changing reality, may be sufficient to generate much of the complexity we perceive as intelligent behavior.

And, perhaps more importantly, they offer a way to fill an architectural gap.

Something need not first decide what deserves intelligent attention.

Millions of possibilities can simply wait upon reality.

Most will amount to nothing.

Some will converge.

And we will call the ones that do intelligent.

If valid, that provides a plausible computational architecture that cognitive science may evaluate, refine, or reject through future empirical investigation.