Is a landscape always continuous?

No. Schedules, board configurations, and other discrete states also form landscapes.

Can changing the neighborhood change local optima?

Yes. A state may be best among one-step moves but not among larger allowed changes.

Does finding a higher peak prove business value?

No. It proves improvement in the chosen objective; the objective must represent the intended priorities.

AI Glossary / Graphs and symbolic AI / Intermediate

State-space landscape

Definition

A state-space landscape is a way to understand an optimization problem by treating candidate states as locations and their objective values as heights. The allowed moves determine which locations are neighbors. Peaks, valleys, and flat regions help explain why a local search method can improve a solution quickly yet fail to find the best solution overall.

What it means

The landscape combines three ingredients: the set of candidate states, their scores, and the neighborhood relation. A picture showing only scores misses a crucial part of the problem because a search method can move only between permitted neighbors.
When higher scores are better, the aim is a global maximum. When scores represent costs, lower is better and the aim is a global minimum. A landscape may have many dimensions or consist of discrete configurations, so the familiar mountain illustration is a teaching device rather than a literal map.

How it works

Define the candidate representation and an objective function, then specify what counts as a nearby alternative. A local search procedure evaluates some neighbors and decides where to move.
A strict local maximum is better than every immediate neighbor, even if a better state exists elsewhere. A flat region contains connected states with equal scores; it may have an improving exit or may form a flat local maximum. The effect of these features depends on the algorithm. A method that accepts only strict improvements behaves differently from one that allows sideways moves, occasional worsening moves, or restarts. Changing the allowed neighborhood can change which states are local optima without changing any scores.

Three practical examples

A staff-scheduling model can score complete schedules and define neighbors by swapping two assignments.
A parameter-tuning exercise can plot model quality across two selected settings to inspect nearby improvements.
A puzzle optimizer can treat the number of violated constraints as a cost and examine configurations reached by one allowed move.

Think of it this way

Think of walking through hills in fog while seeing only nearby ground. Climbing the steepest visible slope can lead to a hilltop that is lower than a distant mountain. The analogy is limited because real optimization neighborhoods need not resemble geographic distance, and high-dimensional landscapes cannot be fully shown as a two-dimensional drawing.

Worked example

Consider an original discrete landscape with states A through H arranged in a line. Only adjacent letters are neighbors. Their scores are A = 1, B = 4, C = 3, D = 5, E = 5, F = 6, G = 9, and H = 7. Higher is better.
Start at A and accept only moves to a strictly higher-scoring neighbor. The search moves from A to B. At B, its neighbors are A with score 1 and C with score 3, so it stops. B is a strict local maximum with score 4. It is not the global maximum: G has score 9.
Starting at D exposes a different difficulty. Its neighbors C and E have scores 3 and 5. Neither is strictly better than D's score 5, so the same rule stops immediately. However, a sideways move from D to E would expose F with score 6, followed by G with score 9. The connected flat region D - E has an uphill exit; it is a shoulder rather than a flat region with no improving exit.
Now change the neighborhood so that a move can jump up to two letter positions. B can directly reach D, improving from 4 to 5. D can then reach F, improving to 6, and F can reach G, improving to 9. B is no longer a local maximum under this new move rule.
No score changed. The change in behavior comes entirely from the neighborhood. This is why a claim that an objective has a particular local optimum should specify the permitted moves or the relevant continuous neighborhood.
Finally, replace every score by its negative and minimize. The same desirable states remain desirable: G becomes the global minimum at −9. The terminology changes from peaks to valleys, but the underlying ordering and search difficulty do not disappear.

Illustrative case study

In a fictional scheduling project, a team improves schedules by swapping one pair of assignments at a time. Repeated runs stop at different scores. The team investigates whether useful changes require coordinated exchanges involving more people.
It experiments with a larger neighborhood and measures both solution quality and the additional evaluation work. It also checks that every proposed schedule remains feasible. The fictional case illustrates how the landscape depends on the move design, without claiming that a larger neighborhood always produces a practical improvement.

Common misconceptions and mistakes

A local optimum is not automatically a poor solution; it may be useful, or it may also be globally optimal. The issue is what the available evidence proves.
Do not assume that a plateau has no exit, or that every local-search method must stop there. Do not confuse a smooth-looking projection with the full landscape: hidden dimensions and constraints can change the picture. Finally, a better objective score is only as meaningful as the objective itself. Missing business constraints can make a mathematical peak undesirable in practice.

Business use

The landscape idea helps explain why optimization results depend on starting points, move rules, and stopping conditions. It helps explain why search becomes stuck.
Teams can compare several starts, inspect recurring bottlenecks, and test alternative neighborhoods within a fixed evaluation budget. Report the objective and feasibility checks alongside the score. When no global guarantee exists, describe the best solution found under the stated search conditions rather than presenting it as the best possible result.

Check your understanding

Why does strict improvement stop at D even though G is reachable through E and F?

Answer: The first move D - E has equal score, so that rule refuses it before discovering the later improvements.

Related terms

Local search
Objective function
Neighborhood
Hill climbing
Simulated annealing
Local optimum
Global optimum
Plateau
Random restart

Learn next

Next, change E's score from five to four. The original path from D toward G now requires an initial worsening move. Explain why merely allowing equal-score moves is no longer enough, and identify which other search mechanisms could explore beyond that barrier.

Sources

Primary source: Russell and Norvig, AIMA, Local Search Algorithms and Optimization Problems, including the landscape discussion and hill-climbing procedure: https://aima.cs.berkeley.edu/4th-ed/pdfs/newchap04.pdf . The eight-state landscape and neighborhood comparison are original teaching examples.

Previous: State space graph Next: State space modeling