Must the teacher be human?

No; a capable automated system can provide demonstrations.

Is behavioral cloning the whole field?

No; it is one important approach.

Can imitation guarantee better performance than the teacher?

No. Such improvement requires additional reasoning and evidence.

AI Glossary / Graphs and symbolic AI / Advanced

Imitation learning

Definition

Imitation learning is a family of methods that learn how to perform a task from demonstrations of another agent's behavior. Instead of specifying every decision rule, a teacher supplies examples of what to do. The learner uses these examples to develop a policy for selecting actions in relevant situations.

What it means

The teacher may be a person, a programmed controller, or another system. Demonstrations can contain observations, actions, and complete trajectories. Their usefulness depends on whether the learner can access the information needed to reproduce the decisions.
Imitation does not necessarily mean replaying an identical recording. A learned policy should respond to the situation it encounters. Nor does copying behavior establish that the learner understands the teacher's intentions or that those intentions are desirable.
The scope is broader than one algorithm. Behavioral cloning directly learns an observation-to-action relationship. Other approaches infer a reward or objective from demonstrations and use it to obtain behavior.

How it works

First define the task, the learner's observations, and its available actions. Collect demonstrations that link the relevant observations with the teacher's choices. In behavioral cloning, train a predictor using these pairs, with a loss suitable for discrete actions or continuous controls.
Sequential deployment creates a complication: the learner's actions influence its next observations. A small mistake can lead to situations absent from the teacher's original demonstrations. Good prediction on recorded expert states therefore need not translate into reliable independent operation.
Dataset aggregation, commonly called DAgger, addresses this by collecting states visited under a learner-influenced policy, requesting the expert's action for those states, adding those labels to the accumulated dataset, and retraining. Its formal guarantees require stated assumptions; the name alone does not promise flawless performance.

Three practical examples

A packing robot learns placement motions from operator demonstrations. Training examples include different object positions. Evaluation checks completed placements, not merely similarity between recorded and predicted motor commands.
A simulated character learns movement decisions from a skilled controller. After a wrong turn, it may need recovery behavior that never appeared in the controller's successful routes.
A document-workflow assistant learns which review step an experienced operator selects next. Its observations must include the document status available at that moment; later approval information would create a misleading training shortcut.

Think of it this way

Think of learning a route by watching a guide. Following the guide's decisions is useful while you stay on the demonstrated route. If you miss a turning, the next useful lesson is how to recover from your actual location, rather than another replay of the original journey.

Worked example

Consider an invented simulation in which a learner must make 20 decisions to complete a route. Assume, purely for this calculation, that each decision independently matches the required action with probability 0.98. Also assume that any mismatch makes the route unsuccessful and that there are no other causes of failure.
The probability of a completely successful route is 0.98 raised to the power 20, approximately 0.6676. Therefore, 98% action accuracy corresponds to about 66.8% route success under these assumptions. The probability of at least one mismatch is about 33.2%.
The expected number of mismatches is 20 multiplied by 0.02, or 0.4. That is not the probability of a failed route: a failed route can contain several mismatches, so counting errors and counting failed routes answer different questions.
Now suppose another learner matches each required action with probability 0.995 under the same assumptions. Its route success probability is 0.995 raised to the power 20, approximately 0.9046, or 90.5%. Improving action accuracy by 1.5 percentage points changes sequence success substantially in this constructed example.
These calculations do not describe the general behavior of imitation learning. Real decisions need not be independent, some mismatches are harmless, and some errors can be corrected. Moreover, action accuracy measured on expert recordings may differ from accuracy in states the learner reaches itself.
To examine that last issue, run the learner in the simulator and retain observations from wrong-turn states. Ask the teacher for the appropriate recovery action at each retained observation. Add these labeled examples to training. Then evaluate complete routes on separate scenarios. Collecting labels does not, by itself, prove that retraining improved performance; that requires the new evaluation.

Illustrative case study

Fictional Willow Packing trains a placement assistant from successful operator demonstrations. Its report emphasizes high agreement with recorded commands, but tests reveal failures after objects shift slightly during handling.
The team discovers that the demonstrations contain almost no recovery situations. Operators naturally corrected alignment before recording, so the dataset omits conditions the learner encounters during its own attempts.
Willow adds controlled recovery demonstrations and separately evaluates unseen object arrangements. Its report now distinguishes command agreement, completed placements, intervention frequency, and recovery outcomes. It also records which observations were available to the model when each decision was made.
The change improves the evaluation design. Whether the revised model improves actual packing performance remains an empirical question, not a result established by this fictional case.

Common misconceptions and mistakes

“Imitation means exact playback.” Policies can generalize across observations.
“The expert's behavior is automatically optimal.” Demonstrations can contain limitations and errors.
“High action accuracy guarantees task success.” Sequence outcomes need separate measurement.
“More demonstrations always solve the problem.” Missing observations or relevant situations can remain unresolved.
“DAgger removes every deployment risk.” Its algorithmic guarantees are conditional, and implementation still needs evaluation.

Business use

Use imitation learning when useful demonstrations are available and explaining every decision explicitly would be difficult. Before collecting data, specify the task outcome and the information that the deployed learner will actually observe.
Budget for demonstration review, recovery examples, and independent task evaluation. Keep entire trajectories or scenarios separate where appropriate when forming training and test sets; adjacent frames from the same demonstration can make a test deceptively easy.

Check your understanding

Why is 0.4 not the route-failure probability?

Answer: It is an expected count of mismatches.

Why collect recovery examples?

Answer: They cover situations absent from successful expert routes.

Why exclude future approval status?

Answer: The deployed assistant cannot use information unavailable at decision time.

Related terms

Policy
DAgger
Inverse reinforcement learning
Trajectory
Behavioral cloning
learning from demonstration
distribution shift

Learn next

Compare action agreement with complete-task success on a small simulation. Then inspect failures to distinguish insufficient demonstrations, missing observations, and mistakes that push the learner into unfamiliar states.

Sources

AIMA glossary, imitation learning: https://github.com/aimacode/aima-glossary/blob/master/glossary.md Ross, Gordon and Bagnell, DAgger research paper: https://proceedings.mlr.press/v15/ross11a/ross11a.pdf Osa and colleagues, An Algorithmic Perspective on Imitation Learning: https://arxiv.org/pdf/1811.06711

Previous: Imieliński - Lipski algebra Next: Immutable Data Types