1. Welcome Welcome to the Dagstuhl seminar on Plan Recognition
Please upload titles for the talks you want to give
We would like everyone to have an opportunity to give a short talk
We have some panel ideas, but these are open to reconsideration – contact me
We will be scheduling incrementally
Scheduled through tomorrow…
Schedules will be re-posted as updated…

2. Panel ideas Should there be a plan recognition competition?
Rational versus fallible agents
Activity recognition, behavior recognition, plan recognition, goal recognition
Oh, my!
Full and partial observability
Generative versus plan library approaches

3. Schedule: Monday AM: Welcome and survey PM:
Jerry Hobbs: discourse and plan recognition
Short talks
George Ferguson
Matthew Stone
Chris Baker: plan recognition and psychology
Panel: a plan recognition competition?
Evening: get acquainted event

4. Schedule: Tuesday AM: PM: Kathy Laskey: probabilistic methods for PR
Short talks
Froduald Kabanza
Francis Bisson
Gita Sukthankar
PM:
Tom Dietterich: learning and plan recognition
David Pattison
Nate Blaylock
Panel: Rational versus fallible agents?

5. Plan Recognition Historical Survey
Henry Kautz
University of Rochester
Robert P. Goldman
SIFT, LLC
Old school plan recognition…
Dagstuhl, April 2011

6. Outline Dimensions of the plan recognition problem
Historical survey of methods
Challenges

7. Dimensions of Plan recognition

8. Keyhole, intended and adversarial plan recognition
Observer non-intrusively watches the agent
Determine how an agent’s actions contribute to achieving possible or stipulated goals
Model
World
Agent’s beliefs

9. Keyhole, intended and adversarial plan recognition
Intended recognition
Agent acts in order to signal his beliefs and desires to other agents
Speech acts – inform, request, …
Discourse conventions
“The 3:15 train to Windsor?”
“Gate 10” [Allen & Perrault]
Symbolic actions
The Statue of Liberty
9/11?
The agent may require a model of the observer.

10. Keyhole, intended and adversarial plan recognition
Agent acts in order manipulate the observer
Deception, bluffing, misdirection, etc. …
Agent and observer will need sophisticated models of each other’s inferences

11. Ideal versus fallible agents
Mistaken beliefs
John drives to Reagan, but flight leaves from Dulles.
The doctor bleeds the patient to cure disease.
Cognitive errors
Distracted by the radio, John drives past the exit.
Jill schedules a doctor’s appointment during her office hours.
Irrationality
John furiously blows his horn at the car in front of him.

12. Output of plan recognition
Activity recognition
Simply identify a known behavior pattern
Goals
Recognize the objective, but not the specific recipes used
Plans
Next action the agent will take?
Best action to aid or counter the agent?

13. Output of plan recognition: likelihood
Most likely interpretation?
Distribution over plans and goals?
The above have subtly different strengths and weaknesses…
Most critical plan or goal?

14. Richness of plans Are actions atomic?
Or do they have parameters?
Structure (e.g., cases)?
Do plans have structure and parameters?
Coreference?
The patient of the plan will be the destination of step one and the patient of step two…
Are there plan libraries at all?

15. Other dimensions Reliable versus unreliable observations
“There’s a 80% chance John drove to Dulles.”
Open versus closed worlds
Fixed plan library?
Fixed set of goals?
Fixed set of entities?
Metric versus non-metric time
John enters a restaurant and leaves 1 hour later.
John enters a restaurant and leaves 5 minutes later.
Single versus multiple ongoing plans
“White knights”
Static versus evolving set of intentions
Abandoning goals: I was going to drive to the store, but the weather was too bad.
Reacting to opportunities: I was going by the playroom on the way from the laundry, so I picked up the toys.

16. Dimensions Relation to agent Model of agent Goals and plans
Observ-ations
Infer
Model (library)
Produce
Intended
(possibly) Irrational
Static
Noisy
Activity
Incomplete
“The Answer”
Keyhole
Partial knowledge
Partial
Goal
Best answer
Adversar-ial
Homo Econom-icus
Dynamic
Complete
Plan
Distribu-tion
Next action

17. METHODS

18. Earliest work Generally in service of language understanding
Often narrative understanding
Understanding indirect speech acts
Allen & Perrault, “Analyzing Intention in Utterances,” AI, 1980
Rich vein of work using plan recognition in dialog understanding and IUI
Will be hearing more from George Ferguson later today!
Methodologically: Mostly shared early enthusiasm for rule-based systems

19. Hypothesize & Revise The Plan Recognition Problem C. Schmidt, 1978
Based on psychological theories of human narrative understanding
Mention of objects suggest hypothesis
Pursue single hypothesis until matching fails
The Plan Recognition Problem C. Schmidt, 1978
Related work from Yale AI Lab: Cullingford’s Script Applier Mechanism, Wilensky’s PAM, etc., 1978
Charniak, Ms. Malaprop, 1978 – Frame-based and used TMS

20. Closed-world reasoning
Infers the minimum set(s) of independent plans that entail the observations
Observations may be incomplete
Infallible agent
Complete plan library
Limited to pasta preparation
Flying Spaghetti Monster
A Formal Theory of Plan Recognition and its Implementation Henry Kautz, 1991

21. Parsing
Vilain use parsing results to characterize computational complexity of plan recognition
There were earlier attempts to parse plans
Parsing techniques closely related to Closed-world reasoning (Built on Kautz and Allen)
Find an explanation that covers all of the observations
Parsing techniques deal poorly with partial ordering, worse with interleaving
Leads to:
Later work on stochastic parsing (Pynadath and Wellman)
Attempts to exploit exotic parsing techniques (Geib)

22. Abduction Reason from effect to cause (C.S. Peirce) People:
Explanation
Diagnosis
People:
Charniak
Hobbs et al., TACITUS
Leads to interest in Bayes nets

23. Bayes Nets DAG-structured models of probability distributions
Came into the fore for diagnostic applications
Challenge: Static Bayes nets for complex domains can be extremely large
Sprinkler
Raining
Grass
wet

24. Bayes Nets
Knowledge Based Model Construction: Dynamically build Bayes nets showing how plans explain actions
Multiple goals
Abstraction hierarchies
Equality reasoning for coreference
Poor treatment of time
“Jack went to the liquor store.”
Was he shopping?
“A Bayesian Theory of Plan Recognition,” Charniak and Goldman, AIJ, 1993.
“Interpretation as Abduction,” Hobbs, Stickel, Martin & Edwards, Proc. ACL, 1988.

25. More on Bayes net methods
Laskey and her colleagues have worked on military domains
Further developed KBMC techniques (e.g. query completeness); coreference, identity uncertainty
Many related techniques
E.g., Hobbs et al. cost-based abduction
ATMSes (d’Ambrosio, Provan, Charniak & Goldman)
Horn logic (Poole)

26. Pending sets
Explicitly models the agent’s “plan agenda” using Poole’s “probabilistic Horn abduction” rules
Bridge between Bayes net and HMM frameworks
Handles multiple concurrent interleaved plans & negative evidence
Number of different possible pending sets can grow exponentially
Happen(X,T+1)  Pending(P,T), X in P, Pick(X,P,T+1).
Pending(P’,T+1) Pending(P,T), Leaves(L), Progress(L, P, P’, T+1).
A new model of plan recognition. Goldman, Geib, and Miller,1999
“A probabilistic plan recognition algorithm based on plan tree grammars,” Geib and Goldman, AIJ, 2009.

27. Version Space Algebra
Recognizes novel plans
Complete observations
A sound and fast goal recognizer Lesh & Etzioni, IJCAI 1995
Programming by Demonstration Using Version Space Algebra Lau, Wolfman, Domingos, Weld.
Related to later work on plan-recognition through planning

28. CHALLENGES

29. Evaluation Ground truth Prediction tasks Difficult to get labeled data
Epistemic question --- do our proposed labelings correspond to any real ground truth?
Prediction tasks
Next action?
Future action?
Good choice of assistive action? Countermeasure?
Can prediction act as proxy for ground truth?

30. Epistemic question
What is the status of the recipes that we postulate as explanations for actions?
Are they taken as being real in some sense?
Corresponding to mental contents?
Identified regularities that really exist in the world?
Data structures that just exist for our convenience

31. Computational difficulties
Computational complexity
Theoretical results
Practical results
Challenges from domains
Some domains inherently ambiguous
Adversarial reasoning
Do we need game-theoretic reasoning
Cooperative as well as adversarial

32. Plan libraries Engineered? Learned? Something in between?
Learned ones often seem impoverished
Engineering seems impossible!

33. Learning Structural learning Parameter learning
Learn the contents of plan libraries (in one form or another)
Parameter learning
Adjust parameters of known libraries
Both offer challenges related to those of evaluation
Plan recognition may be done in service of learning, as well as the other way around.
Infer goals to learn novel recipes

34. Imperfections Imperfect agents Imperfect observations Imperfect models
Imperfect information
Imperfect reasoning
Imperfect task performance
Challenging for non-empirical algorithms
Imperfect observations
Imperfect models
Including seemingly-irrelevant actions

35. User models
In many domains, the behaviors exhibited are not just a function of the actions, goals and plans, but agent characteristics, as well.
Developing clean ways to combine agent-dependent and – independent information is a challenge going forward.
Often per-agent training is unacceptable.

36. Sensing In many cases it is difficult to sense the agents’ actions:
Labeling actions in primitive sensor data
Vision
Network packets
Linguistic utterances
Hardware/software hybrid systems
E.g., oil refinery --- user can go out and use a wrench un-observed
Conventional software
Even Horvitz et al. report difficulties “seeing” actions of Microsoft Office users
Mixed streams
Individual actions in network packet streams

37. Coreference and quantification
In some domains we don’t have object identity and permanence and the number of agents simply handed to us.
Story understanding
Military situation interpretation
Identity hypotheses enter into plan recognition

38. Anomaly detection
Often appealed to as a solution for detecting some phenomenon that is difficult to model:
Intrusion behavior in computer security
Terrorist behavior in tracking and camera data
Dementia-induced behavior in tracking elderly subjects
Accuracy requires deep understanding of the models’ properties
Stationarity (often violated in computer security)
“Size” and “shape” of normal behaviors
As always, it’s hard to get something for nothing.

39. The Role of State
Many (but not all) plan recognition systems represent only the state of the planning agent.
The state of the environment is modeled implicitly, if at all.

40. Groups Teamwork Role recognition
Friendly: recognize teammates’ intentions to coordinate and aid
Hostile: recognize opponents’ intentions to hinder and obstruct
Role recognition

41. Hypothesis retrieval
Some early work assumed that there were enough candidate hypotheses that retrieval could be an issue

42. Predictive and explanatory inference
A lot of concern in early work about combining top-down and bottom-up inference

43. Actions with weak diagnostic power
E.g., computer security
We would desperately like to know the attacker’s motivations
But what do we do with
Get access to the target
Gain administrator privileges on the target…

44. Coffee and then henry’s turn…