|
|
|
Raja Das, Jim Hanson, Jeff Kephart, Gerry
Tesauro |
|
IBM Thomas J. Watson Research Center |
|
Hawthorne, NY 10532, USA |
|
|
|
Collaborators: |
|
Steven Gjerstad, Jason Shachat (IBM Research) |
|
Jonathan Bredin (Dartmouth), Weng-Keen Wong (CMU) |
|
|
|
|
Introduction |
|
|
|
Experimental setup |
|
|
|
Agent architecture and bidding strategies |
|
|
|
Results |
|
|
|
Speculation |
|
|
|
|
|
|
|
|
|
|
Agents will play important role in dynamic
pricing |
|
Automated, dynamic posted pricing (buy.com) |
|
Bidding services (eSnipe) |
|
Automated businesses will automate their pricing |
|
e-utilities, Application Service Providers, Web
Services |
|
|
|
Automated dynamic pricing is useful when |
|
Many prices to set |
|
Frequent or quick pricing decisions |
|
Agent performance rivals or exceeds that of
humans |
|
|
|
|
|
Goals |
|
Show agents can outperform humans in pricing
scenario |
|
Understand collective agent-human dynamics |
|
|
|
Another chapter in Human vs. Machine tradition |
|
Backgammon (Tesauro: TD-gammon) |
|
Chess (Campbell et al.: Deep Blue) |
|
Checkers (Schaeffer: Chinook) |
|
|
|
But some important differences |
|
More players |
|
Incomplete information (e.g., other players’
valuations) |
|
Independent, asynchronous moves |
|
|
|
|
|
|
|
|
|
|
|
|
|
Human subjects |
|
recruited from local colleges and IBM Research |
|
given interactive instructions and test |
|
paid in proportion to surplus |
|
Setup |
|
6 Humans, 6 Agents |
|
6 Buyers, 6 Sellers |
|
Each agent shares limit prices with a human |
|
Experiment |
|
9 to 16 3-minute periods |
|
Limit prices change every 3-5 periods |
|
Record bids, asks, trades |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
Goal: Achieve market equilibrium with minimally
intelligent traders (Cliff 1997) |
|
|
|
Idea: Adjust bid/ask price based on market
activity |
|
Adjust bid/ask price towards new trade prices |
|
Includes a momentum term |
|
If a competitor’s bid not taken, adjust offer to
beat competitor’s bid |
|
|
|
We modified ZIP to handle |
|
Continuous time |
|
Multiple units with different profit margins |
|
|
|
|
|
|
|
Use market history to estimate belief function
f(p): |
|
probability of bid/ask at price p being accepted |
|
fb(p) = [TBL(p) + AL(p)] / [TBL(p) +
AL(p) + RBG (p)] |
|
|
|
Choose p that maximizes E[surplus] |
|
Sb(p) = fb(p) * (Limit –
p) |
|
|
|
We modified GD to handle |
|
Persistent order queue |
|
Pathological collective behavior |
|
Strategizing over multiple units |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
Agents won by substantial margins in all
experiments |
|
~20% more surplus than novice humans |
|
~5-7% more surplus than experienced humans |
|
|
|
Agents and humans interact with one another |
|
Not two decoupled markets |
|
~30-50% of trades are agent-human |
|
|
|
Previously unobserved dynamical effects |
|
Persistent far-from-equilibrium trades |
|
Enhanced scalloping effect |
|
|
|
|
|
Agents perform substantially better than humans
in CDA |
|
Better decisions |
|
Faster execution |
|
|
|
Considerable room for improving bots |
|
Stronger CDA bidding strategies will be
developed |
|
|
|
Agents will supplant humans in CDA & other
auctions |
|
Agents in the trading pit |
|
But humans still guiding their behavior |
|
Agent-human hybrids will continue to co-evolve |
|
|
|
|
|
|
|
|
|
Notes