David Pardoe

1
TacTex-05 A Champion Supply Chain Management
Agent

• David Pardoe
• Peter Stone

The University of Texas at Austin Department of
Computer Sciences
2
Supply Chain Management

• Research goal automate the process
• Trading Agent Competition (TAC SCM)
• Many challenges
• TacTex-05 (2005 winner) - agent composed of
  several interacting components
• prediction
• optimization
• adaptation

3
Outline

• Summary of TAC SCM
• TacTex-05 agent design
• Adaptive aspects of TacTex-05
• Competition results and experiments
• Conclusion

4
TAC SCM

• Agents compete as manufacturers
• 220 simulated days per game (15s each)

5
Component Procurement

• Suppliers production capacity fluctuates
• Prices depend on suppliers free capacity

6
Customer Negotiation

• 16 computer types in 3 segments
• Daily number of RFQs fluctuates

7
Factory Scheduling

• Limited production capacity
• Daily storage cost for all inventory

8
(No Transcript)
9
Outline

• Summary of TAC SCM
• TacTex-05 agent design
• Adaptive aspects of TacTex-05
• Competition results and experiments
• Conclusion

10
(No Transcript)
11
Demand Model

• Goal predict future customer demand
• Bayesian approach adapted from DeepMaize
  (Kiekintveld et al. 2004)

12
Order Probability Predictor

• Want to predict P(order offer price)
• Linear predictor for each computer type

13
Demand Manager

• Given resources and predictions, determine
• production schedule
• deliveries
• offers on all of todays RFQs
• All done with greedy scheduling algorithm

14
(No Transcript)
15
(No Transcript)
16
(No Transcript)
17
Supplier Model

• Estimate each suppliers free capacity from
  offers
• Use estimates to predict future offer prices

18
Supply Manager What to Order

• Goal maintain a threshold inventory

19
Supply Manager When to Order

• Given a desired delivery, when to send RFQ?
• Assume todays price pattern holds

20
Outline

• Summary of TAC SCM
• TacTex-05 agent design
• Adaptive aspects of TacTex-05
• Competition results and experiments
• Conclusion

21
Adaptation

• Different opponents lead to different situations
• Adapt by modifying predictions
• Make use of game logs

22
Two Areas of Adaptation

• Initial orders and endgame sales
• Important, but difficult to reason about
• Agents may handle as special cases
• Update predictions during these periods

23
Outline

• Summary of TAC SCM
• TacTex-05 agent design
• Adaptive aspects of TacTex-05
• Competition results and experiments
• Conclusion

24
Final Results

• Adaptation important
• ordered 95,000 components on first day
• SouthamptonSCM 22,000 Mertacor 18,000

25
(No Transcript)
26
Experiments

• Experiments analyzing agent components
• Use TAC Agent Repository
• Compare modified versions of TacTex-05
• Test adaptation against different opponents

27
Results

• Start-game adaptation
• competition results very atypical
• End-game adaptation
• beats fixed strategies in experiments
• Predictive models
• supplier price predictions most important
• Often better to wait to order components
• tradeoff price vs demand certainty

28
Outline

• Summary of TAC SCM
• TacTex-05 agent design
• Adaptive aspects of TacTex-05
• Competition results and experiments
• Conclusion

29
Related Work

• Many TAC SCM agent descriptions
• SouthamptonSCM He et al. 2006
• Mertacor Kontogounis et al. 2006
• DeepMaize Kiekintveld et al. 2006
• CMieux Benisch et al. 2006
• Available from TAC website http//www.sics.se/tac

30
TAC News

• 2006 TAC SCM competition complete
• Won by TacTex-06
• Most important addition use learning to predict
  future changes in computer prices
• TAC in 2007 3 games
• TAC Classic
• TAC SCM
• New market design game

31
Conclusion

• Introduced TAC SCM
• Described TacTex-05
• prediction
• optimization
• adaptation
• Future work
• additional learning, adaptation
• focus on component price prediction, ordering

32
Thank You!