Hypothesis number grows exponentially. (Geib and Goldam '09) ... B. C. Is plan A part of the correct hypothesis? No, it
Interactive Plan Recognition (Formerly Sequential Plan Recognition) Reuth Mirsky, Roni Stern, Ya’akov (Kobi) Gal, Meir Kalech Department of information systems engineering Ben-Gurion University of the Negev
1
Plan recognition “Infer the plan of the agent given observations and a data base of recipes” Kautz (1986)
2
Hypothesis representation • Plan = tree
• Hypothesis = set of plans
• Other representations (e.g. Ramírez and Geffner ‘09)
3
Hypothesis representation The output of a PR algorithm is a set of hypotheses
Mix all
Pair-wise Mix
4
Number of hypotheses Hypothesis number grows exponentially (Geib and Goldam ‘09) 100000 Hypotheses
10000 1000
VirtualLabs TinkerPlots Simulated
100 10 1 1
2
3
4
5
6
7
# observations
8
9
10
5
The disambiguation problem
6
Prior work • Pruning • Kabanza and Filion. Controlling the hypothesis space in probabilistic plan recognition. IJCAI. 2013. • Wiseman and Shieber. Discriminatively reranking abductive proofs for plan recognition. Twenty-Fourth International Conference on Automated Planning and Scheduling. 2014. • Mirsky, Gal, Shieber. CRADLE: an online plan recognition algorithm for exploratory domains. [under submission].
• Less expressive representation • Avrahami-Zilberbrand and Kaminka. Incorporating observer biases in keyhole plan recognition (efficiently!). AAAI. Vol. 7. 2007. • Geib. Delaying Commitment in Plan Recognition Using Combinatory Categorial Grammars. IJCAI. 2009. • Sukthankar and Sycara. Hypothesis pruning and ranking for large plan recognition problems. AAAI. Vol. 8. 2008.
7
Contributions • The Interactive Plan Recognition Process (IPRP) • Soundness and Completeness proof • Policies for solving IPRP • Extensive empirical evaluation
8 Mirsky, Gal, Stern and Kalech. Sequential Plan Recognition. AAMAS (Extended Abstract). 2016. Mirsky, Stern, Gal and Kalech. Sequential Plan Recognition. IJCAI. 2016.
IPRP Iterative process for disambiguating the hypothesis space Coming up: •Which hypotheses to discard? •Which plan to choose?
9
The Refinement Criteria A Create Flasks
B
Perform Test Mix All
Perform Test
Write Results
C
Solve Problem Create Flasks
Solve Problem
Write Results
Create Flasks
Solve Problem Perform Test Pair-wise Mix
Write Results
10
The Refinement Criteria Is plan A part of the correct hypothesis?
A Create Flasks
B
Perform Test Mix All
Perform Test
Write Results
C
Solve Problem Create Flasks
Solve Problem
Write Results
Create Flasks
Solve Problem Perform Test Pair-wise Mix
Write Results
11
The Refinement Criteria Is plan A part of the correct hypothesis? A Yes, it is part of my plan
Then B and C might also be part of your plan B
Create Flasks
Perform Test Mix All
Perform Test
Write Results
C
Solve Problem Create Flasks
Solve Problem
Write Results
Create Flasks
Solve Problem Perform Test Pair-wise Mix
Write Results
12
The Refinement Criteria Is plan A part of the correct hypothesis? A No, it is not part of my plan Then neither B nor C can be part of your plan B
Create Flasks
Perform Test Mix All
Perform Test
Write Results
C
Solve Problem Create Flasks
Solve Problem
Write Results
Create Flasks
Solve Problem Perform Test Pair-wise Mix
Write Results
13
The Refinement Criteria Is plan B part of the correct hypothesis? A Create Flasks
B
Perform Test Mix All
Perform Test
Write Results
C
Solve Problem Create Flasks
Solve Problem
Write Results
Create Flasks
Solve Problem Perform Test Pair-wise Mix
Write Results
14
The Refinement Criteria Is plan B part of the correct hypothesis? A No, it is not part of my plan Then B cannot be part of your plan, but C might
B
Create Flasks
Perform Test Mix All
Perform Test
Write Results
C
Solve Problem Create Flasks
Solve Problem
Write Results
Create Flasks
Solve Problem Perform Test Pair-wise Mix
Write Results
15
The Refinement Criteria Is plan B part of the correct hypothesis? A Yes, it is part of my plan We Cannot discard hypotheses with A
B
Create Flasks
Perform Test Mix All
Perform Test
Write Results
C
Solve Problem Create Flasks
Solve Problem
Write Results
Create Flasks
Solve Problem Perform Test Pair-wise Mix
Write Results
16
The Match Criteria B
C
Solve Problem Create Flasks
Perform Test
Write Results
Solve Problem
Create Flasks
Mix All
Create All
A
Solve Problem Create Flasks
Perform Test
Create All
Mix All
Write Results
Perform Test
Write Results
17
The Match Criteria B
C
Solve Problem Create Flasks
Perform Test
Write Results
Solve Problem
Create Flasks
Mix All
Create All
Is plan B part of the correct hypothesis? No, Plan B is not a part of my plan Then A and B can not be part of the correct hypothesis
A
Solve Problem Create Flasks
Perform Test
Create All
Mix All
Write Results
Perform Test
Write Results
18
The Match Criteria B
C
Solve Problem Create Flasks
Perform Test
Write Results
Solve Problem
Create Flasks
Mix All
Create All
Is plan B part of the correct hypothesis? Yes, Plan B is part of my plan Then all plans can still be part of your plan
A
Solve Problem Create Flasks
Perform Test
Create All
Mix All
Write Results
Perform Test
Write Results
19
Soundness and completeness • Completeness: IPR does not remove any hypothesis that can be refined to the correct hypothesis • Soundness: Every hypothesis IPR keeps can be refined to the correct hypothesis All Hypotheses
Removed Remaining
Correct
20
Which plan to query next? • Random: choose a random plan • Most Probable Hypothesis (MPH): choose a plan from the most probable hypothesis
• Most Probable Plan (MPP): choose the most probable plan
• Entropy: choose the most informative plan 21
Reduction in Hypotheses by Strategy
22
Discussion • After 1 query • 39% reduction in hypotheses (simulated domain) • 46% reduction in hypotheses (VirtualLabs domain)
• Entropy strategy outperforms all other strategies
• Find an optimal querying policy • Apply IPRP in real world domains • Open source code on github:
https://github.com/ReuthMirsky/IPR
[email protected]
23