Probabilistic Plan Recognition Using Off-the-Shelf Classical Planners

Source: DOI Authors: Miguel Ramírez, Hector Geffner Date: 2010

Core idea

Observed actions should update a probability distribution over candidate goals rather than collapse immediately into one asserted intention.

Key claims

• Goal inference is model-relative and evidence-relative.
• Multiple candidate goals should remain live when evidence is ambiguous.
• Confidence should come from a clear inference method rather than from impressionistic certainty.

Harness takeaway

A multiplayer harness should store inferred goals as competing, revisionable hypotheses with evidence links and explicit confidence semantics, not as durable facts about a person.