Does this require machine learning?

No. A controller can use hand-designed rules, analytical models, or learned policies.

Is every internal memory value part of the state representation?

Only if the chosen representation needs it to predict relevant outcomes; abstraction is often necessary.

Does metalevel reasoning mean self-awareness?

No. It concerns decisions about computation and information processing, without implying subjective experience.

AI Glossary / Time series and signal processing / Advanced

Metalevel state space

Definition

A metalevel state space is a representation of the possible internal computational states of an agent or program. Its states describe what the program has already computed, what alternatives remain, and what information or resources are available for further reasoning. Transitions correspond to computational actions, such as expanding a search node, evaluating an option, or stopping to act.

What it means

An agent faces two kinds of choices: what to do in its task environment, and what reasoning to perform before deciding. The object level concerns the task itself, such as choosing a route. The metalevel concerns the decision process, such as which route estimate to refine next.
Two programs can face the same physical situation but occupy different metalevel states because they have explored different alternatives. This term describes computational organization, not consciousness or an agent's subjective awareness.

How it works

Define a computational state containing information relevant to future reasoning decisions. It may include a partial search tree, evaluated actions, current best solution, uncertainty about alternatives, elapsed computation, and remaining resources. Define available computation actions and how their possible results update that state.
A metalevel controller compares the expected benefit of further computation with its cost and the value of acting now. In dynamic settings, physical state and deadlines may also need to be included because time spent computing can change the task environment. A useful abstraction must preserve information needed for these decisions without reproducing every low-level machine detail.

Three practical examples

A game-playing program chooses which branch of a partially explored game tree to examine next before committing to a move.
A route planner decides whether to refine another candidate route or return the best route already found within its response budget.
A scheduling system allocates computation among alternative solution methods and stops when further expected improvement no longer justifies the delay.

Think of it this way

Think of a researcher deciding which experiment to run before choosing a design. The designs belong to the original decision problem; the notebook of completed tests, remaining uncertainties, and available time describes the reasoning situation. A metalevel state captures that second perspective in a form that a controller can use.

Worked example

Consider an original one-step metareasoning example with costs measured in a common utility-equivalent unit. Option A has known task cost 10. Option B has unknown task cost: 6 or 18 with equal probability. Without more computation, B's expected cost is 12, so a risk-neutral agent chooses A at cost 10.
A computation can reveal B's cost perfectly and costs one unit. If B is revealed as 6, choose B; if it is 18, choose A. Expected task cost after the information is 0.5 × 6 + 0.5 × 10 = 8. Including computation gives total expected cost 9, an improvement of one compared with acting immediately.
The expected value of this information before its cost is 10 − 8 = 2. If the same computation instead costs three, total expected cost becomes 11, so acting immediately is preferable among these two choices.
Before computation, the metalevel state includes A's known cost, uncertainty about B, and the computation option. After computation, it includes the revealed B value. The object-level choice still concerns A versus B. This result assumes perfect information, no other computations, risk-neutral expected cost, and no environmental change during the computation; it is not a general optimal-search policy.

Illustrative case study

Illustrative fictional case: Acorn Scheduling runs several improvement procedures on a delivery plan. The controller tracks the incumbent plan, which neighborhoods have been searched, and the remaining response time. It chooses another search step when the expected improvement justifies the delay.
The team evaluates total decision quality, including computational cost, rather than celebrating a better route found after the dispatch deadline. The fictional case illustrates why reasoning choices can be modeled explicitly without assuming that unlimited computation is valuable.

Common misconceptions and mistakes

The metalevel state space is not simply the physical task state with a more impressive name. A partial search tree and a truck's actual location describe different things, even when both matter to an overall controller.
Do not assume that metareasoning is free or automatically optimal. Deciding which computation to perform is itself a potentially difficult problem, and its state space can be much larger than the object-level space. A greedy value-of-computation rule can also miss benefits from sequences of computations that are useful only in combination.

Business use

Metalevel representations can help allocate limited computation across planning, search, and decision-support tasks. They make tradeoffs between answer quality, latency, memory, and processing cost explicit.
Specify what counts as a computation, how its cost is measured, and what evidence estimates its expected benefit. Test stopping decisions and resource allocation on representative tasks. When the environment changes while the program reasons, include the cost of delay and stale information rather than evaluating computation only by its eventual answer quality.

Check your understanding

What changes after the revealing computation in the example?

Answer: The agent's computational information about B, while the available task options remain A and B.

Why is the information worth two units before cost?

Answer: It reduces expected task cost from ten to eight.

Why reject it when it costs three?

Answer: The total expected cost becomes eleven, worse than acting at cost ten.

Would the calculation remain valid if delay changed both task costs?

Answer: Not without extending the model to include that change.

Related terms

Metareasoning
Object-level state space
Anytime algorithm
Bounded rationality
Stopping rule
computational state
value of computation
search control
metalevel policy

Learn next

represent a small search program by its explored nodes, frontier, and incumbent solution. Compare a fixed computation budget with an explicit stopping rule, then examine cases where two computations together are more useful than either alone.

Stopping rule

Sources

Sources: AIMA official index, metalevel state space: https://aima.cs.berkeley.edu/aima-index.html . Stuart Russell, research overview on bounded optimality and rational metareasoning: https://aima.eecs.berkeley.edu/~russell/research-bo.html . Nicholas Hay, Principles of Metalevel Control, indexed thesis introduction: https://aima.cs.berkeley.edu/~russell/papers/hay-phdthesis.pdf . Primary indexed sources consulted. Example and case are original.

Previous: Metalevel learning Next: Metamerism