Must the most likely state have probability above 50%?

No. It only needs to have probability at least as large as every other individual state.

Can there be more than one most likely state?

Yes. A tie requires a stated handling rule or retention of multiple candidates.

Is planning with the most likely state always wrong?

No. It can be a useful approximation, but its suitability depends on uncertainty, action consequences, and the task's requirements.

AI Glossary / Graphs and symbolic AI / Intermediate

Most likely state

Definition

The most likely state is a state assigned the greatest probability by an agent's current probability distribution over possible states. In robotics and partially observed planning, it is often used as a single best estimate on which to base a simpler plan. Selecting that estimate does not remove the uncertainty represented by the rest of the distribution.

What it means

A robot may be uncertain about its location, an object's position, or whether a route is blocked. Its belief state represents several possibilities with associated probabilities. Extracting the most likely state reduces that distribution to one candidate, which can simplify planning.
The phrase means most probable under the specified model and evidence, not known to be true. The leading state may have probability below one-half if the alternatives are numerous. Several states may tie. In continuous settings, a mode maximizes a probability density rather than the probability of an exact point, so the representation and parameterization require additional care.

How it works

For a finite state set, compare the probabilities in the current belief distribution and select a maximizing state. A most-likely-state planning approach plans as though that state were the actual one, executes some action, receives observations, updates the belief, and may replan.
The approximation can save computation, but it discards information relevant to action choice. A less probable state may make a proposed action costly or dangerous. Basic goal-directed planning on one estimate may also overlook the value of gathering information. More elaborate approaches can retain the belief distribution or explicitly switch to information-gathering behavior.

Three practical examples

A localization system can report its highest-probability grid cell while preserving the full location distribution for downstream decisions.
A manipulation planner can use the most likely object pose to propose a motion, then evaluate whether remaining pose uncertainty makes that proposal unsuitable.
A monitoring dashboard can show a leading operational-state estimate alongside alternatives, avoiding the impression that the displayed label is certain.

Think of it this way

Think of choosing the leading candidate in a forecast while keeping the other candidates on the page. The leader summarizes the ranking but does not inherit all their probability. The analogy is limited because a robot's state can be high-dimensional, and the best action may depend more on the consequences of being wrong than on which candidate ranks first.

Worked example

Consider an invented robot belief over three possible locations: A has probability 0.45, B has 0.35, and C has 0.20. A is the most likely individual state, even though the probability that the robot is somewhere else is 0.55.
Suppose a shortcut action has cost 2 in A, 20 in B, and 20 in C. A conservative route costs 8 in every state. Planning only for A selects the shortcut because 2 is less than 8.
Using the complete belief and minimizing expected cost gives the shortcut cost 0.45(2) + 0.35(20) + 0.20(20) = 0.9 + 7 + 4 = 11.9. The conservative route has expected cost 8, so that criterion prefers it. The example demonstrates that choosing the most likely state and choosing the best action are different operations. Costs and probabilities are invented; no real robot safety assessment is implied.

Illustrative case study

In a fictional inspection-robot prototype, two corridors look similar to a camera. The localization display shows the more probable corridor, and an early planner treats that display as certain. A test reveals that a motion suitable in one corridor is undesirable in the other.
The team preserves both location hypotheses and checks proposed actions against the current belief. In ambiguous cases it considers an additional observation before proceeding. Engineers then test the revised behavior under realistic sensing errors. The case illustrates a design issue with collapsing uncertainty; it does not claim that an extra observation always resolves ambiguity or that a specific robot was deployed.

Common misconceptions and mistakes

Do not confuse the most likely state with the mean state. For two separated location hypotheses, an average may lie in an implausible place. Nor is a most likely current state the same as the most likely entire sequence of states; temporal decoding is a different query.
Avoid interpreting the largest probability as automatically high confidence. A leading probability of 0.45 is still compatible with substantial uncertainty. Also do not label the estimate “maximum likelihood” without checking the calculation: maximizing a posterior state distribution is generally a maximum-a-posteriori estimate and includes prior information.

Business use

A single state estimate can make interfaces and planning systems simpler, but its presentation should match the decision risk. For a low-cost display, the leading label may be adequate. For an irreversible action, omitted alternatives can matter even when their probabilities are small.
Define when a planner may use the simplified estimate, what uncertainty information remains available, and when another observation or a belief-aware decision is needed. Evaluate decisions, not just state-label accuracy. A system can often select the correct leading label yet still make poor choices if the occasional errors have disproportionate consequences.

Check your understanding

Change the worked belief to probabilities 0.80, 0.10, and 0.10 while keeping the costs fixed. What does expected-cost comparison select?

Answer: The shortcut costs 0.8(2) + 0.1(20) + 0.1(20) = 5.6 in expectation, below the conservative route's 8. The action changes because the full probability distribution changes, not because A newly becomes the leading state.

Related terms

Learn next

Calculate the probability threshold for A at which the shortcut and conservative route have equal expected cost, when all non-A states cost 20. Solve 2p + 20(1-p) = 8 to obtain p = 2/3. Then compare that decision threshold with the much weaker condition that A merely rank first.

Sources

Sources: Hsiao, Lozano-Pérez and Kaelbling, Robust Belief-Based Execution of Manipulation Programs, original paper describing and extending the most-likely-state approach: https://lis.csail.mit.edu/pubs/tlp/wafr08.pdf . Russell and Norvig, original textbook robotics passage, indexed excerpt: https://people.engr.tamu.edu/guni/csce642/files/AI_Russell_Norvig.pdf . All numerical costs and the robot case are original illustrations.

Previous: Most general unifier Next: Most probable explanation