Presentation on theme: "4/17: Plan Recognition Several slides borrowed from Kautzs overview talk."— Presentation transcript:
4/17: Plan Recognition Several slides borrowed from Kautzs overview talk
Points to make.. Activity vs. Plan recognition– in activity recog, you are setting up the HMM yourself. In Plan recognition (as in Goldman), the Bayes net is automatically set up. Symbolic ones just find consistent plans--but even consistency makes no sense if you can make errors Set minimization loses the likelihood.. Does it matter whether the underlying domain is deterministic or probabilistic? Observing actions or state?
..More points to make The idea of plan recognition becomes an oxymoron if you dont have HTNs or Goal Schemas Hongs goal schemas are really primitive HTNs.. (cant handle round trip) Kautz talks about Infallible agent assumptionbut what about the infallible recognizer assumption (which means the recognizer has the correct model of the agents actions?) Sometimes reasonable to assume actions will be given as input (instead of state) e.g. unix domain; game domains etc Whether you need a single most likely plan or multiple plans depends on what you really need to do with the plans once recgonizedintrusion detection; prefetching.. –Write the applications slide –The woman and old man joke.. Talk next class about State Estimation problem (or put it in the context of action vs. state input). Why are particle filters useful in plan recognition (as against localization of oneselfwhere the agent knows the actions it is doing..) Write a whole slide on the importance of goalsgoals are not just any old set of states.
Plan/Goal Recognition Recognize the plan and/or higher level goal(s) of the agent from partial observations of the agents behavior (actions done; states of the world visited) –Recognition can be used to either aid or thwart the agents plans and/or make sense of their actions Very active area of late with many shades.. –Intention Recognition; Goal Recognition; Plan recognition; Activity Recognition; Behavior Recognition Applications include: –Dialog Understanding (if you recognize an agents plan/goals) you can do a better job of understanding their actions –Activity/Behavior recognition (recognize suspicious activities; recognize need to assist cognitively impaired people) –Intrusion detection/avoidance –User intent recognition to provide intelligent user interfaces –Assistance (if you recognize what the agent is trying to do, you can help the agent do it better/faster etc.) But if you mis-recognize, you will have the woman and old man on the Bus scenario.
Levels of Recognition Physical movement Movement sensor fires Behaviors Running, grasping, lifting, … Plans Getting a drink of water Describes conventional way of achieving a goal Goals Quench thirst Research efforts normally may start and end in the middle… e.g. from plans goals or From movement behaviors Generally, the PR literature separates Goals from normal state variables (can think of goals as higher-order state variables… Thus the popularity of hierarchical models)
On the importance of pre-set goals in plan recognition In normal planning settings, we take the view that any set of states can be designated goal states. So, with n state variables, we can have 2^2^n possible goal states. In plan recognition, we have to assume that the agent is interested in a much smaller set of goals (than 2^2^n). If you allow all sets of states to be potentially goal states, then there is nothing to recognize (whatever state you see the agent reaching is potentially the state the agent wanted to reach!) The set of potential agent goals is circumscribed in two ways: Provide hierarchical plan libraries (e.g. HTN schemata). You recognize the plan under observation by parsing it in terms of the HTN schema [e.g. Kautzs original work; Charniak/Goldman work etc.] Provide (hierarchical) goal libraries. You match the states encountered with the possible goals they are satisfying. (e.g. the Hong work on goal graphs; most of the work on activity recognition).
Planning vs. Plan Recognition It is intuitively clear that a plan recognizer can benefit from the knowledge of goals of the agent, as well as the domain model –But which domain model? The *real* domain model? The model according to the agent acting The model according to the recognizer The that the recognizer thinks the agent has of the world? –In theory, all of these could be different.. If agent thinks that speaking loudly will help blind people understand better, then the agent may be rational w.r.t. its model) If the recognizer thinks that when people step out the front door, they just stand behind that door until they decide to come in, then that is the model it will use –Plan recognition is an inherently multi-agent activity! Planning is also relevant in two ways –If the recognizer knows the agents goal and the domain model, then it can figure out what the agents policy should be (notice that the agent may not be optimal even with respect to its potentially faulty modelso the plan that the recognizer comes up with may only be an approximation to what the agent actually does) –If the recognizer wants to help or thwart the agent, it will need to make a plan based on its recognition of the agents goals as well as the preconditions of the agents actions You can facilitate or defeat the preconditions based on whether you want to help or thwart
Interactive Spelling Correction --and word completion Given a dictionary of words A partial or complete typing of a word Complete/Correct the word argmax c P(c|w) argmax c P(w|c) P(c) / P(w) P(w|c) Error model –What is the probability that you will type w when you meant c? –Different kinds of errors (e.g. letter swapping) have different prob –Consider edit distance P(c) language model -How frequent is c in the language that is used? In Auto-completion, you are trying to suggest most likely completion of the word you are typing…(even in face of typing errorsa la Ipod.)
Comparing Spelling correction to plan recognition… Dictionary of words ~ Library of Plans Errors ~ wrong plans –Is the passenger looking furtively around while going through security using a wrong normal travel plan or right blow up the plane and not get caught plan? Language Model ~ Plan/Goal Corpus.. (with what goals are more likely)
Dimensions of the behavior recognition problem Keyhole versus interactive Keyhole Determine how an agents actions contribute to achieving possible or stipulated goals No model of the observer – fly on the wall Interactive Actions performed by an agent to signal to another agent Speech acts Model social conventions & agents models of other agents
Dimensions of the plan recognition problem Ideal versus fallible agents Mistaken beliefs John drives to Reagan, but flight leaves Dulles. Cognitive errors Distracted by the radio, John drives past the exit. Irrationality John furiously blows his horn at the car in front of him.
Dimensions of the plan recognition problem Reliable versus unreliable observations Theres a 80% chance John drove to Dulles. Open versus closed worlds Fixed plan library? Fixed set of goals? 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 Also, does the PR agent observe actions or the state? Normally, only the state is observed. In computer game worlds etc where the agent initiates an action by pressing specific buttons, actions can also be observed
Dimensions of the plan recognition problem Desired output: Set of consistent plans or goals? Most likely plan or goal? Most critical plan or goal? Interventions observer should perform to aid or hinder the agent? Also, is the recognition done at the end of the observed behavior, or online. In intrusion detection, you may want to know all possible consistent plans (so you can defeat all of them); in passive assistance, you may want to look at a set of likely plans (so you can prefetch information to speedup their execution). In active assistance, you will likely need to stick to a single most likely plan/goal
Approaches to plan recognition Consistency-based Hypothesize & revise Closed-world reasoning Version spaces Probabilistic Stochastic grammars Pending sets Dynamic Bayes nets Layered hidden Markov models Policy recognition Hierarchical hidden semi-Markov models Dynamic probabilistic relational models Example application: Assisted Cognition Can be complementary.. First pick the consistent plans, and check which of them is most likely (tricky if the agent can make errors)
Hypothesize & Revise The Plan Recognition Problem C. Schmidt, 1978 The Plan Recognition Problem Based on psychological theories of human narrative understanding Mention of objects suggest hypothesis Pursue single hypothesis until matching fails
Closed-world reasoning A Formal Theory of Plan Recognition and its Implementation Henry Kautz, 1991 A Formal Theory of Plan Recognition and its Implementation Infers the minimum set(s) of independent plans that entail the observations Observations may be incomplete Infallible agent Complete plan library Similar to Bottom-up HTN parsing..
Bottom-up HTN Parsing vs. Plan Recognition One way of doing HTN planning we discussed involved keeping incremental parses of the plans being generated by the underlying non primitive planner [Barrett & Weld; 1994] In a way the HTN parser was recognizing what the searcher below was doing However, it only cared to make sure that there is a non-empty set of partial parses We Goal model &put the applications up front..
Goal Graph-based recognition Separately identified Goal schemas At each level, state literals supporting goal schemas are identified An action done is relevant to a goal schema if its effect supports it directly or indirectly The consistent goal schemas are those for which all the observed actions are relevant. Output htese…. [Hong; JAIR 2000] Cant recognize sequential goals unless goal schemas are in CTL
Version Space Algebra A sound and fast goal recognizer Lesh & Etzioni A sound and fast goal recognizer Programming by Demonstration Using Version Space Algebra Lau, Wolfman, Domingos, Weld. Programming by Demonstration Using Version Space Algebra Recognizes novel plans Complete observations Sensitive to noise
Approaches to plan recognition Consistency-based Hypothesize & revise Closed-world reasoning Version spaces Probabilistic Stochastic grammars Pending sets Dynamic Bayes nets Layered hidden Markov models Policy recognition Hierarchical hidden semi-Markov models Dynamic probabilistic relational models Example application: Assisted Cognition Can be complementary.. First pick the consistent plans, and check which of them is most likely (tricky if the agent can make errors) Next class