Imitation Learning

There's a shortcut past trial-and-error: learn from someone who already knows. But copying an expert is subtler than it looks, and the way it fails is deeply instructive.

When a skilled practitioner already makes good decisions, you don't need to rediscover good behaviour through costly exploration — you can learn it from their example. Imitation learning builds a policy from demonstrations rather than from a reward signal or a model. It's the formal version of apprenticeship, and it's how you might turn the hard-won judgement of your best project and risk experts into a reusable, teachable policy.

1Behavioural cloning, and its hidden trap

The obvious approach, behavioural cloning, treats imitation as ordinary supervised learning: collect examples of the situations the expert faced and the actions they took, then fit a policy that reproduces those actions. Simple and often a fine start. But it hides a serious flaw. The learner is only ever trained on the situations the expert visited — and experts, being good, mostly stay on well-managed paths. The moment the learner makes a small error and drifts somewhere the expert never went, it has no guidance, so it errs again, drifting further into unfamiliar territory. Small mistakes compound into large ones. This is distribution shift, and it is the central difficulty of imitation.

Figure 1. The learner tracks the expert until a small error pushes it off the demonstrated path. Off that path there are no examples to imitate, so errors compound and the gap widens. This is why a playbook copied action-for-action fails in situations it never covered.

2Two better answers

Query the expert where the learner actually goes. Instead of a fixed set of demonstrations, iteratively run the learner, note the situations it drifts into, and ask the expert what they would do there — then add those to the training set. Over rounds, the learner gets guidance precisely for the situations it's prone to reach, and the compounding-error problem is tamed.
Infer the goal, not the moves. Inverse reinforcement learning asks a deeper question: rather than copying the expert's actions, work out what objective the expert appears to be pursuing, then optimise that. Recovering the underlying intent generalises far better than mimicking surface behaviour — a policy that understands the goal can handle situations the demonstrations never showed, because it knows what it's trying to achieve.

Key idea

Copying what an expert does breaks down the instant you leave the paths they walked. Capturing what the expert is trying to achieve survives into new situations — which is the difference between a brittle checklist and genuine transferred judgement.

Where Part III leaves us

Part III faced an unknown model: whether to explore or exploit, and whether to learn the model (model-based), learn actions directly (model-free), or learn from an expert (imitation). But every method in Parts I–III assumed you could at least see the current state — you knew which situation you were in. Real projects deny even that: status is reported late, partially, and sometimes wrongly. Part IV takes on the hardest case — deciding well when you cannot directly observe the true state at all.

What it means in practice

To capture expert judgement, start by recording what your best practitioners do — but know that a demonstration set alone yields a brittle policy that fails in exactly the off-script situations where judgement matters most. Strengthen it by asking your experts specifically about the awkward, unfamiliar cases a junior would stumble into, and by drawing out the objectives behind their choices, not just the choices. The most transferable thing you can extract from an expert is not their moves but their aims.

Handbook application: from concept to controlled practice

Purpose. This expanded section turns the original page into a practical handbook. It preserves the supplied material and adds a repeatable way to apply, check and review Imitation Learning. It does not replace a contract, legislation, a controlled standard, competent engineering judgement or specialist advice.

The operating aim is to convert the subject into a governed decision, owned work, usable evidence and a reviewable outcome. Read the original explanation first, then use the workflow and checks below to convert knowledge into evidence.

Use Imitation Learning as a decision instrument rather than an administrative form. The subject terms—learning, behavioural, cloning, imitation, expert—need an explicit connection to the project objective, business value and stakeholder commitments. Before completing the artefact, write one sentence stating who will use it, what decision it supports and when that decision is required.

Apply a disciplined information model. Separate facts supported by evidence, forecasts derived from a method, assumptions awaiting validation, constraints that limit choice, risks that may occur, issues that already exist and actions assigned to people. Each material entry should have an owner, date, status and next review point. Where probability or impact scores are used, define the scale so different reviewers interpret it consistently.

A baseline is useful only when changes are visible. Give the artefact an identifier, version, approval state and effective date. Define which changes require reapproval, how superseded versions are retained and where supporting evidence is stored. During reviews, focus on exceptions, decisions and trends rather than reading every field aloud. Record the decision and rationale, not merely that a meeting occurred.

Close the loop beyond delivery. Confirm acceptance criteria, unresolved items, transferred responsibilities and operational ownership. Where benefits are expected, identify the outcome measure, baseline, target, observation period and owner who remains accountable after the project team disbands. Lessons should describe the condition, consequence and reusable action; a generic statement such as “communicate better” cannot improve the next project.

Step-by-step operating method

Clarify the decision. Name the outcome, sponsor, affected stakeholders and decision that this work must enable.
Set boundaries. Record scope, assumptions, constraints, dependencies, tolerances and escalation conditions.
Plan the evidence. Define deliverables, measures, owners, due dates and acceptance criteria before execution.
Control delivery. Compare actual performance with the baseline, assess changes and manage risks and issues explicitly.
Close the loop. Confirm acceptance, transfer ownership, capture lessons and track benefits beyond handover.

Completion and governance protocol

Start with a short drafting workshop involving the accountable owner and the people who hold the evidence. Complete high-consequence fields first: objective, scope, owner, baseline, acceptance, dependencies and escalation. Mark unknowns as assumptions or actions rather than hiding them behind vague prose. Circulate a review draft, resolve conflicting interpretations, baseline the approved version and place the next review date in an owned schedule.

Information type	Minimum useful content	Review test
Outcome	Observable change and intended recipient	Not merely a deliverable or activity
Measure	Definition, baseline, target, frequency and source	Two reviewers would calculate it the same way
Ownership	One accountable role plus contributors and approver	Authority matches responsibility
Uncertainty	Assumption, risk or issue with response and trigger	Status reflects current reality
Control	Version, approval, review date and change rule	Current baseline is identifiable

Common failure modes and recovery actions

1. Watch for

Producing a document with no named decision or accountable owner.

Recovery: Return to the governing definition or requirement and restate the decision in one sentence.

2. Watch for

Mixing risks, current issues, assumptions and actions in one unstructured list.

Recovery: Separate evidence from assumption, assign an owner and set a date for validation.

3. Watch for

Measuring activity or output while leaving the intended outcome undefined.

Recovery: Run a small counterexample, boundary test, pilot or independent check before proceeding.

4. Watch for

Accepting changes without evaluating effects on value, scope, schedule, cost and risk.

Recovery: Record the consequence, decision and rationale, then update the controlled baseline.

5. Watch for

Closing the project at delivery even though benefit ownership has not transferred.

Recovery: Escalate when the issue affects safety, compliance, acceptance, material value or an agreed tolerance.

Review checklist

Which decision or commitment does this artefact support?
Who owns each action, risk, acceptance and post-project benefit?
What is the baseline and what variance triggers escalation?
Where is the evidence that the result was accepted and transferred?
Are mandatory requirements distinguished from recommendations and illustrative values?
Are sources, assumptions, units, dates and versions recorded closely enough to reproduce the decision?
Have safety, legal, ethical, stakeholder and operational consequences been considered at the appropriate level?
Is there a named owner and a trigger for review, escalation, change or retirement?

Questions for deeper application

What is the most important distinction a practitioner must preserve when applying Imitation Learning?