eric horvitz - microsoft.com€¦ · capture real-world data on location and mobility digital...
TRANSCRIPT
Eric Horvitz
Joint work with John Krumm and Adam Sadilek
Behavioral
data store
Population activity in the wild
Paradigm for Digital Experimentation
World
simulator
Behavioral
data store
Population activity in the wild
World
simulator
What-if studies
Paradigm for Digital Experimentation
Behavioral
data store
Population activity in the wild
World
simulator
What-if studies
Paradigm for Digital Experimentation
Behavioral
data store
Population activity in the wild
World
simulator
Studies of incentives & responses
What-if studies
Paradigm for Digital Experimentation
Behavioral
data store
Population activity in the wild
World
simulator
Studies of incentives & responses
Realized design
What-if studies
Paradigm for Digital Experimentation
Decomposition Feasibility Estimation
Human
intellect
Machine
intellect
Availability
Competence
Cost
Interpreter Tasks
Solver
Planner Solution
E. Kamar, S. Hacker, E. Horvitz. Combining Human and Machine Intelligence in Large-scale Crowdsourcing, 2012.
Coordination
(AI, Human)
HUMAN
HUMAN HUMAN
HUMAN
AI
AI
AI AI
AI
HUMAN HUMAN
HUMAN
HUMAN
AI AI AI
HUMAN
HUMAN
AI
HUMAN
Coordination
(AI, Human)
REPAIR
(i)
REPAIR
(i)
TRANS
(j,k)
MT MT MT
TRANS
(i,j)
TRANS
(i,j)
Example: Lingua Mechanica
D. Shahaf, & E. Horvitz. Generalized Task Markets, AAAI 2010.
Search &
Optimization
TRANS
(k,l)
D. Shahaf, & E. Horvitz. Generalized Task Markets, AAAI 2010.
D. Shahaf, & E. Horvitz. Generalized Task Markets, AAAI 2010.
D. Shahaf, & E. Horvitz. Generalized Task Markets, AAAI 2010.
$
Interpreter Tasks
Solver
Planner Solution Perception Action
Human
intellect
Execution Coordination Composition
Machine
intellect
Sensors
Effectors
D. Shahaf, & E. Horvitz. Generalized Task Markets, AAAI 2010.
People Cameras Cranes Forklifts Cars ….etc.
Capture real-world data on location and mobility
Digital experimentation via parameter sweeps
Seek insights on shaping “fabric” of collaboration
Preferences, incentives and links to collaborative fabric
Implications & directions
A. Sadilek, J. Krumm, E. Horvitz. Crowdphysics: Planned & Opportunistic Crowdsourcing for Physical Tasks.
Actively “shaped” routing graph: set proximity d, dwell time t
Canonical Crowd Physics Problem
A. Sadilek, J. Krumm, E. Horvitz. Crowdphysics: Planned & Opportunistic Crowdsourcing for Physical Tasks.
Studies
Locations via GPS-coded tweets
Sweep through d,t
d
Shaped Contact Graphs
t
d
$ d
Contact Graphs from Tweets
A. Sadilek, J. Krumm, E. Horvitz. Crowdphysics: Planned & Opportunistic Crowdsourcing for Physical Tasks.
Actively “shaped” routing graph: set proximity d, dwell time t
Shaped Contact Graphs
d
$ d d ’
d ’
Contact Graphs from Tweets
A. Sadilek, J. Krumm, E. Horvitz. Crowdphysics: Planned & Opportunistic Crowdsourcing for Physical Tasks.
Actively “shaped” routing graph: set proximity d, dwell time t
Shaped Contact Graphs
d =100 meters t =30 minutes
21 days Cities: 60x60 km
Contact Graphs from Tweets
A. Sadilek, J. Krumm, E. Horvitz. Crowdphysics: Planned & Opportunistic Crowdsourcing for Physical Tasks.
Shaped contact graphs: Experiments with design
Actively “shaped” routing graph: set proximity d, dwell time t
Canonical Task: Routing Physical Objects
Synchronization & sequencing of effort
Local vs. centralized plans
Canonical task: Route packages
• Coverage & reachability
• Efficiency (delivery time)
• Optimality of routing heuristics
• Robustness to graph ablation
• Locale sensitivity
Random infinite graphs
Static, homogeneous networks, repeat lattice structure
Studies: Properties of finite, real-world contact graphs
Leverage for task coordination
Lévy flight: random walk w/
heavy-tailed probability distribution.
Frequent short-hops, rare long jumps (Mandelbrot)
P(x) a x-b b = 2.2
Small-world phenomenon?
Graph diameter O(log n)
Shortest path discoverable?
Global: Dijkstra, Floyd-Warshall
Assume: Complete, static knowledge
Real-time: Uncertainty about future locations
Local opportunistic routing
Use heuristics based on historical data
60 x 60 km bounding box around Seattle, 6 months
450 x 450 m cells, start & destination (ci, cj )
Test: Final 35 hours
Train: Rest of 6 mos.
Local policy:
> For each location l, consider statistics of proximity to l. > Keep package or handoff based on historical proximity
Global: Dijkstra
Phase transition @
d =350 meters t =30 minutes
Understanding & Shaping Collaborative Systems
Studies of crowd physics for sensing & acting in world
Design, optimization of collaborative substrate
Rich area, multiple directions for enhancing collaboration
Opportunities include new approaches to epidemiology,
e.g., where reachability & efficiency are minimized