University of Rochester
Activities
Abductive Inference of Multi-Agent Interaction
Capture the Flag Data Collection
Representing Beliefs & Goals of Multiple Agents
Modal Markov Logic
Recognizing Indoor Activities using Multi-Modal
Data
Fusing RFID and Machine Vision
DEPARTMENT of COMPUTER SCIENCE
Multi-Agent Interaction
Many agent behaviors can only be
understood in the context of the
actions of other agents
Exercising?
Being chased?
Chasing someone?
Location alone provides a surprisingly rich source of
information about behavior
GPS data can be used to learn a probabilistic model of a
individual’s common activities (Liao, Fox, & Kautz 2007)
Goal: learn models of groups and interactive activities
Relational learning problem: ideal for ML
Needed: dataset of competitive & cooperative interaction
DEPARTMENT of COMPUTER SCIENCE
Capture the Flag
Capture the Flag Data Collection
UR campus
Up to 150 x 300 m area
Complex topology
14 players, 8 games
GPS loggers
Accuracy varies 1-9 m
Average game 12 m
DEPARTMENT of COMPUTER SCIENCE
Start of Game
DEPARTMENT of COMPUTER SCIENCE
End of Game
1.
2.
3.
4.
red & orange
guarded by
green
green leaves
prisoners
violet releases
red & orange
red captures
flag
DEPARTMENT of COMPUTER SCIENCE
Ground Truth
For supervised learning
methods, need to create a
labeled training set
First attempt: record voice
annotations from players
Failed: players too involved
to accurately comment on
their actions
Second attempt: post-hoc
annotation
Created general annotation
tool for relations over GPS
streams
DEPARTMENT of COMPUTER SCIENCE
Supervised Weight Learning
Discrete features calculated from GPS streams
Supervised learning applied to simple 2-slice model
Precision: 46% (second by second)
Recall 64%
12 hours to label 1 hour of training data
DEPARTMENT of COMPUTER SCIENCE
Observations
Humans can accurately perceive interactive
behaviors
High agreement between annotators
GPS noise often obscures geometric details
Reasoning about intention over extended
temporal context disambiguates action
DEPARTMENT of COMPUTER SCIENCE
Year 2 Goals
Improve quality of data (features) using
physical constraints
Hard constraints: walls
Soft constraints: paths
CRF “snapping” tool
Model long temporal dependencies
Unsupervised learning: discover behaviors,
tactics, strategies
DEPARTMENT of COMPUTER SCIENCE
Representing Beliefs & Goals
of Multiple Agents
Abduction often requires reasoning about the
establishment of “propositional attitudes”
Belief, desire, intention, commitment, …
Example: principles of communication:
If A tells B that P, then A believes P.
If A tells B that P, then B will believe that A wants B
to believe P.
If A is cooperative with B, and B wants P, then A
will want P.
Such principles are defeasible
DEPARTMENT of COMPUTER SCIENCE
Modal Operators
In logic
Predicates relate one object to another
Modal operators relate objects (agents) to
propositions (sentences)
Different modalities can be axiomatically
characterized
Deductive closure:
Transitivity:
B(a, P) B(a, P Q) B(a,Q)
B(a, P) B(a, B(a, P))
DEPARTMENT of COMPUTER SCIENCE
Modal Operators in Markov Logic
ML defines a probability distribution over
propositional truth assignments
Idea: define probability distribution over
assignments that are modally consistent
Non-modal atoms
Modal atoms
Modal consistency check
DEPARTMENT of COMPUTER SCIENCE
Inference
Complexity of consistency check
Depends on target modal logic
Belief (KD45):
Unbounded nesting: PSPACE-complete
Bounded nesting: NP-complete
Modal Markov Logic Inference
Rejection sampling
Optimizations
Cache g(M)
Compute g(M) incrementally
DEPARTMENT of COMPUTER SCIENCE
Year 2 Goals
Implement MML in Alchemy
Applications
Understanding indirect speech acts
Capture the flag
Establishing knowledge by perception
Representing degrees of belief
Functional modal operators
B(a, B(b, P) 0.9) 0.75
DEPARTMENT of COMPUTER SCIENCE