Must a conditional plan be invented during execution?

No. Its branches can be prepared in advance and selected as observations arrive.

Is the cheapest expected plan cheapest in every situation?

No. The worked example has a lower average cost but a higher cost on the large-item branch.

Does an observation always reveal the true state?

No. A model may include noisy or incomplete observations, requiring decisions based on remaining uncertainty.

AI Glossary / Graphs and symbolic AI / Intermediate

Conditional plan

Definition

A conditional plan specifies an action and alternative continuations chosen according to information obtained during execution. Instead of committing to one fixed action sequence, it says what to do after each relevant observation. Conditional plans are useful when an agent does not initially know the state or when action outcomes can differ.

What it means

A plan can be prepared before execution while still adapting during execution. Its branches encode those adaptations. For example, an agent can inspect an item and then follow the continuation appropriate to the inspection result.

The branches must depend on information the agent can actually obtain. Writing if the hidden state is favorable does not create an executable plan unless the agent has an observation or belief-based decision mechanism that supports that choice.

In a partially observable decision process, a finite conditional plan can be represented recursively: perform an action, then map each possible observation to a further conditional plan. This provides a compact conceptual bridge between a decision tree and a policy for acting under uncertainty.

How it works

The planner models possible states, available actions, observations, and the objective. It considers how an initial action can affect the world and what information may become available afterward.

It then constructs a suitable continuation for each relevant observation. During execution, the observed result selects the next branch. A branch can itself contain another action and another observation-dependent choice.

Evaluating a plan requires accounting for the branches that can occur. In a probabilistic setting, expected value combines branch outcomes using their probabilities. Other planning settings may require success under every modeled possibility. A plan that performs well on average is not automatically guaranteed to achieve a goal in every case.

Noisy observations require particular care. The observed label is not necessarily the true state, so the planner may need to update a belief distribution and choose a continuation that reflects remaining uncertainty.

Three practical examples

1. A document-processing agent checks whether an attachment is readable, then follows either an extraction path or a request-for-replacement path.

2. A simulated delivery agent observes which passage is available and selects the corresponding route, provided its plan covers the modeled observation outcomes.

3. A diagnostic workflow performs an information-gathering step before choosing among alternative actions, including the time and cost of obtaining that information.

Think of it this way

Think of a conditional plan as a set of travel directions with explicit junction instructions: follow this route if the displayed sign says one thing, and take the alternative if it says another.

The analogy is limited by observation quality and coverage. A missing or misleading sign may not fit either branch. The plan needs an appropriate response to such possibilities if they are part of the environment model.

Worked example

Consider a toy handling task with two item sizes. An item is small with probability 0.6 and large with probability 0.4. A general handling action costs 6 units and works for either size. A small-item action costs 2 units but works only for small items.

An inspection costs 1 unit, reveals the size perfectly, and does not change it. A conditional plan is: inspect; if the result is small, use the small-item action; if large, use the general action.

The total cost is 3 units on the small branch and 7 on the large branch. Expected cost is 0.6 times 3 plus 0.4 times 7 = 4.6 units.

The fixed plan that always uses the general action costs 6 units, so this conditional plan has lower expected cost under the stated assumptions. It is nevertheless more expensive on the large branch.

If inspection instead costs 3 units, its expected total becomes 6.6. Branching is therefore not automatically beneficial: information-gathering cost matters. These comparisons concern the specified alternatives, not an exhaustive proof of optimality over every possible plan.

Illustrative case study

Fictional case: Poplar Handling uses the toy model to explain a proposed inspection workflow. Its first presentation highlights the cheap small-item branch but omits inspection cost and the more expensive large-item branch.

The reviewers rebuild the plan as an observation tree and calculate each complete path from start to finish. They also record that the example assumes perfect size identification; a fallible inspection would require a different calculation and possibly different continuations.

This case demonstrates how a conditional plan should be assessed. It does not claim observed cost savings or recommend the fictional process for real material handling.

Common misconceptions and mistakes

Do not count only the action after a branch. Earlier actions and information-gathering steps also contribute to the path's cost and outcome.

Do not confuse a conditional plan with a conditional effect. A conditional plan selects future actions; a conditional effect describes how one action changes a state under specified conditions.

Do not assume a plan covers every possibility because it has several branches. Check the observation model and identify unmatched outcomes. Also distinguish a planned contingency from replanning, which computes a new plan after receiving information.

Business use

For business teams, conditional plans make exception handling and information needs reviewable. They show which observations drive decisions, where an additional check is worth its cost, and what happens when the expected path cannot be followed.

A useful review includes complete branch costs, observation reliability, and the objective used to compare alternatives. The number of branches can grow rapidly with planning depth, so practical systems may use compact policies or limited planning horizons. Those implementation choices should preserve the intended response to relevant observations.

Check your understanding

A plan branches on a property that the agent cannot observe or infer from available information. Is that branch instruction executable as written?

Answer: No. The plan needs an available information source or a different decision rule that uses what the agent actually knows.

Related terms

Learn next

Draw a two-step observation tree and evaluate every complete path. Then introduce an imperfect observation model and study belief updates to see why branching on an observed signal differs from branching on the true state.

Sources

Malik and colleagues, An Efficient, Generalized Bellman Update for Cooperative Inverse Reinforcement Learning, POMDP background section: https://aima.cs.berkeley.edu/~russell/papers/icml18-cirl.pdf . The primary paper's action-plus-observation-continuation definition was consulted. The handling-cost example and fictional review are original, with simplifying assumptions stated explicitly.

Previous: Conditional odds Next: Conditional probability