CMSC 477/677 - Spring 2007
Discussion Questions for Class #22, April 19
Reading: Davis and Smith, "Negotiation as a
Metaphor for Distributed Problem Solving," Artificial Intelligence, 1983;
Brooke and Durfee: "Congregating and Market Formation," AAMAS-2002.
Davis and Smith
- Contract nets were originally designed for use in cooperative
distributed problem-solving systems. Would contract nets be useful in
self-interested agent communities? What modifications would need to
be made?
- What problems might arise in a contract net system if not every
agent were able to communicate with every other agent?
- What problems could arise if an agent bid on two different tasks,
and were selected for those tasks, but couldn't actually handle them
simultaneously? One way to prevent this situation would be to forbid
an agent from having outstanding bids on multiple incompatible tasks.
What are the disadvantages of that solution? (Hint: What happens if a
manager takes a long time to evaluate bids, or has some pending bids
and is waiting for bids from other agents?)
- Do you think that "negotiation" is a good term for the
bid-response-select cycle of contract nets? Negotiation in the common
sense generally involves more back-and-forth. What situations might
require more extensive negotiations among agents?
- What are the benefits and disadvantages of these extensions to the
contract net protocol:
- Focused addressing
- Directed contracts
- Availability announcements
- In the distributed sensor system scenario, the top-level monitor
node uses the position information provided by other agents to
partition the coverage of the map into roughly equivalent regions.
The area contractors then do the same to select group contractors
How hard do you think it is to do that partitioning? Does this seem
like a genuinely distributed solution?
- How close do you think the distributed DSS solution will be? One
issue is that some vehicle contractors who send bids to the closest
area manager will not be selected, but cannot then be used by another
area manager. Is this a problem with the contract net concept, or
with the particular implementation described here?
Brooks and Durfee
- Can agents belong to more than one congregation in this work?
How would changing this assumption affect the problem?
- The value functions over goods are defined to be linear. Is this
a good assumption in general? How would changing this assumption
affect the problem and corresponding solution?
- What is an "optimal allocation" in this problem?
- The threshold "tau" separates goods into "sufficiently valuable"
and "insufficiently valuable." How is this threshold used?
Is this a strong enough measure of
value to make the market model realistic? How sensitive do you think
the behavior of the system is to the selection of tau?
- What is the nature of the utility functions U_a? Do arbitrary
utility functions make sense, or are there constraints that would
always (or usually) be observed by reasonable utility functions?
Does the presupposition of such a utility function represent a
significant limitation of this approach?
- How do the agents in this paper choose which congregation to join?
Do you think the methods described by Brooks and Durfee would be
useful for our class competition environment? Why or why not?
- The experimental results show profit for sellers increasing over
time. Is this a good thing for the buyers? Why or why not?
- The caption on Figure 1 indicates that the graph is plotting
cumulative profit per iteration; the label above the graph indicates
average profit per iteration. Are these the same? If not, which one
do you think is intended?
- What is the relationship between Brooks and Durfee's work and that
of Davis and Smith?