Matchmaking

Unstructured Agent Matchmaking
Experiments in Timing and Fuzzy Matching
Elth Ogston and Stamatis Vassiliadis
Computer Engineering Laboratory
TU Delft
Are there elements of coordination within
large multi-agent systems that can be
obtained “for free”? i.e.
•Without complicated agent algorithms
(planning, scheduling, intelligence)
•Without external structure
(facilitators, directories, blackboards, know
topology)
Matchmaking: how do agents that require an
outside service find other agents who are
willing to provide that service?
•Assume redundancy of providers and consumers
(MAS are open, flexible and component based)
•Simple agents
•Coordination without outside help
•How studied? - Simulation of an abstract model.
•Results? - We find that there are conditions under which simple
unaided agents do find matches
•This paper – checking two further conditions, timing and how
matches are determined
Talk Organization
•General philosophy
•Overview of our model and previous results
•Some new results on timing and forms of
matching
•Summary of further work
Philosophy - scalability
Multi-agent systems can in theory be world/internet size.
However they often make use of systems components, like
directories, that don’t scale well… why?
Humans tend to believe in (central) control
(God, aliens, The FBI, Mom)
Scientists and engineers who design computers are trained
to see order in the world.
Philosophy – sloppy systems
Natural systems tend to be redundant and full of failures.
Lets try looking at coordination not as beautifully
interlocking clockwork but as an cloud that just happens to
look like an elephant when you squint a bit, turn it upside
down, and ignore that part over there….
Philosophy – matchmaking
thought experiment
Imagine a number between 1 and 10.
How would you find someone else in the room with the
same number?
•Broadcast
•Broker
•Ask your neighbors
Now scale up, find someone in Madrid with a number
between 1 and 100,000
Model - Components
Agents
2 11
5
7
3
18
Tasks
Categories
Connections
Model -Movement
A
C
2 11
C
11
5
B
5
“Shuffle”
11 5
“Cluster”
2
B
5
11
A
7
7
3
18
18 3
Model - Characteristics
•There are several “good” matches available
•We aren’t looking for the global best match
•Not all agents need to be successful
•No centralized directory
•No predefined structure
•Agents are simple
•Agents only know about their immediate surroundings
Previous Results
• Matches are found
– Limited by the number of task categories and
the number of neighbors to each agent
• Limiting cluster size creates a distributed
system
• Replacing tasks creates a dynamic system
New Results
• System timing doesn’t play a role in
coordination
• Fuzzy probabilistic category matches
produce the same behavior as discreet
deterministic matches
Agents moving in sync vs. agents moving in a
random order
Deterministic matches vs. probabilistic matches
Further Work
• AAMAS 2002 – comparison of a peer-topeer auction with a centralized auction
– P2P shows same auction behavior
– As we add more agents P2P has constant message costs
vs. linear for a central auctioneer
More Info….
http://ce.et.tudelft.nl/~elth/