Summary and discussion Session 3: Using Economics for Resource Control Tuomas Sandholm Washington University Slides: compressed Postscript (1228K)

Summary

Automated Negotiation Systems

• Users delegate only small tasks to agents, ensuring small maximum losses.
• Heterogeneous agents negotiate with each other on behalf of their owners.
• Agents might know rules of a transaction better than their owners, allowing them to strike a better deal.
• In business, firms are constantly reorganizing to be more efficient. Corporations create subsidiaries to take advantage of smaller economies of scale, while smaller companies try to federate. What can agents do to find the right balance?
• With little common structure for agent interaction, there is a need to standardize protocols and strategies.
• It is unreasonable to enumerate all possible strategies - use game theory to calculate the best ones.

Are markets good for resource allocation?

• Agents may sell tasks or groups of tasks. Clustering of tasks may allow better marketing or easier planning.
• Given many vendors selling many different bundles of goods, a problem arises of finding the optimal assignment. There may be many local maximas.
• Further complications arise if a trade can involve more than two parties.

Leveled commitment

• Enforcement of exchange can be a problem.
• Possible solution: set up a reward for contract completion and a penalty for defaulting (for both the buyer and seller).
• This system allows partial commitment and adds flexibility.
• Works fine when agents are cooperative, but what about competitive?
• Does not need synchronization, unlike contingency contracts.
• Provides a unique computable Nash equilibrium.

Unenforced exchange?

• What if transactions between agents are susceptible to failure or mischief?
• Relevant to commerce.
• Using sequenced delivery and payment can be effective if the good can be broken up into pieces.
• Need to establish deadlines for transactions.
• How to avoid network and agent saturation with large amounts of communication?

Discussion

Q:	Is it reasonable to assume that we know probability distributions?
A:	Only leveled commitment requires such knowledge.
Q:	When does negotiation break down? What are the alternatives?
A:	Not everyone has to negotiate with everyone else all the time, so overhead is limited. In addition, mediation can make negotiation efficient (a la NASDAQ).
Q:	How do you model misperceptions?
A:	Don't make assumptions on distributions.
Q:	How do you model the fact that agents might not know about their own failures?
A:	We don't yet.