Strategies for a Intelligent Agent in TAC-SCM 28 th September, 2006 Based on studies of MinneTAC (TAC-SCM 2003)
Quick Overview ● The TAC-SCM game actually consists of 2 separate, but inter-related sub-games. ● One game is played in the the market where the agents have to buy supplies ● Second game is played in the market where agents must sell their finished goods
MinneTAC : Agent Outline ● Component-based architecture (similar to DeepMaize) ● Decision & Responsibilities delegated to components: Raw Materials Manager : Manages Purchases Assembly Manager : Decides what to assemble Sales Manager : What RFQs to respond to, and with what price quotes Since the Sales Manager is the where the actual action starts, we'll look at the strategies for it...
What Strategies Are There? ➢ Customer-Demand Driven (Build-to-Order) ➢ Supply Driven
Customer-Demand Driven ● Environment: Assumes that customer demand decides what & how much to make ● Goal of Sales Manager: Maximize profit on a bagged order (via Raw Materials Manager) ● Immediate Benefit: Flexibility to stop doing business in unprofitable environment
Strategy: Maximize Sales Profit The strategy relies only on details in RFQ to decide the offer price This gives a 6-dimensional Order Probability: OrderProbability = offer_price x quantity x lead_time x reserved_price x penalty x product_type And Profit... Expected Profit = Profit x Probability of acceptance
Supply Driven ● Environment: Assumes what customer demand could be, coupled with decides as per past history of its offers' acceptance what & how much to make ● Goal of Sales Manager: Predict a target acceptance rate as close to the actual acceptance rate ● Immediate Benefit: More dynamic in an even more uninformed market
Strategy: Optimize Sales With Demand The strategy relies on details in RFQ to decide the offer price, and also calculates Acceptance rates and demand estimates This gives a 5-dimensional Order Probability: OrderProbability = offer_price x customer_demand x lead_time x reserved_price x product_type And Target Acceptance Rate... TAR product = (available_inventory) x (products_produced) x (num_of_days_left) Optimistic Demand Estimate
What are the differences? Customer-Driven ● Work on restricted data set ● Tries to sell out its inventory of Finished Goods towards the end ● Doesn't rework price calculations as regularly Supply-Driven ● Work on a more expansive, probabilistic set of data ● Tries to sell out its inventory of Finished Goods from the start ● On basis of target acceptance and actual acceptance rates
What was observed
What was observed...
What Fits Best? Customer-Driven ✔ Profitable in an overall increasing price scenario ✔ Works best if customer demand is not 100% satisfied ✔ Tends to hold on to the finished goods in the inventory till better prices come along ✗ Towards the end, a lot of the inventory may be sold of cheaply Supply-Driven ✔ Adapts rapidly to demand and price fluctuations in the market ✔ Tends to sell finished goods in the inventory rapidly from the start with a pessimistic view, making it more competitive with agents having similar traits ✔ Due to relative low inventory of finished goods, it will also sell of fairly cheaply, bu the cumulative loss incurred for this stage is low ✗ On an overall game play, this fails to make most of the market
Conclusion ● Agent clearly cannot adopt any one strategy alone. Balance is required. ● Knowledge of the nature of competing agents helps ● Estimation of customer-demand can solve the bottle-neck ● Split the strategies between the Raw Materials Mgr and Sales Mgr to share & cooperate on information
Reference Source Strategies for a Sales Component of an Intelligent Agent for TAC-SCM 2003 Elena V. Kryzhnyaya University of Minnesota
Thank You! Kunal Khatua Dept. of Computer Science Univ. of Texas at Austin