TAC-SCM Project Plan



TAC SCM Project Plan

Matthew Olson

CIS 690

William Hsu

Problem Statement

The Trading Agent Competition (TAC) Supply Chain Management (SCM) is a game to model a computer manufacturer. Each agent represents a different computer manufacturer. The agent must acquire the different parts it needs for its production center. The end product is sold as demanded by the market [3].

A final score is determined by how well the agent met its client’s needs. The score is determined by bank account size. Each agent starts out with $0 and can go in the hole. Interest is charged while the agent is in the hole and accrued if it has a positive balance. Of the agent uses his bank account to purchase the parts it needs and is paid for each end product bought [3].

On each turn, the agent must “1) Negotiate supply contracts, 2) Bid for customer orders, 3) Manage daily assembly activities, and 4) Ship completed orders to customers” [3]. This can be seen in the 3 parts of the day shown in the next three screen shots.

[pic][pic][pic]

The item I am interested in studying is predicting demand. There are several factors on demand from a single company. Market-share is one of these factors. I believe, and a notation in the code suggests, that more can be done with this factor.

Background

I found two agents with open source that could be used to be the basis of my program. These two are RationalSCM [2] and MinneTAC [4].

Both programs are written in java, which is good because I am the most familiar with it. The first problem I run into with RationalSCM is that it is not very well organized. I am also finding very little information on its performance and where it is lacking.

MinneTAC, on the other hand, is very well documented. It also points to its weakness, market-share [5]. It was actually assumed that market-share would be 1/6. In the real world, it is never like this. Thus we have a good point to improve on.

I had to intensely study MinneTAC before I could make any changes. My understanding of what is going on is pretty good. On a given day in the game, data is dumped into MinneTAC’s repository by the server. Based on the data, triggers are set off. The data is analyzed and the supplier manager section makes demand predictions. These predictions are then multiplied by market-share to make a decision on what parts to order. The attempt is then made to meet the supplier’s price to get these parts. Basically, offers are put in the repository and are sent back to the server.

Methodology

In order to predict changes in market-share, I will program in a version of Winters’ Method. Winters’ Method is used as a forecasting and smoothing function.

Winters’ Method is given as:

Level: Ei=U(Ei-1+Ti-1)+(1-U)Yi

Trend: Ti=VTi-1+(1-V)(Ei-Ei-1) [1].

The level, E, is set for the purpose of using in the trend. The trend, T, is, in our case, market-share. So, we see that market-share is based on its current value and the difference in levels.

From this we see two constants, U and V. These had to be derived from experimentation. I finally settled on U=V=0.995.

Y is the variable that is observed. In the case of the improved agent, Y=(our orders)/(total orders).

Evaluation Methods

There are a couple of ways we can evaluate the improved agent. First of all, can the improved agent beat the old baseline? The second is by how much are we now winning by? In this way we can see if market-share is a valid way to get better performance.

Results

The first experiment I did I put the improved agent and the baseline agent to 22 runs. Our agent won against the baseline 68.2% of the time. There were also four dummies inserted into each game that had negligible bearing on the results. Money-wise, you see the results below.

[pic]

I then reran the experiment with some variations. This time I copied the improved agent four times and did something different with each one. So we kept the baseline. Agent 0 is the same as previously described. Agent 1, is the result of averaging market-share over 2 days. Agent 2, is the result of averaging market-share over 3 days. Agent 3, is the result of averaging market-share over 4 days. In Agent 4, I halved market-share for the first 22 days to try and stimulate growth those days. The results are shown below.

[pic]

Milestones

09/30/2006: project plan (first draft)

10/07/2006: project plan

11/18/2006: coding of agent

11/23/2006: debugging and testing of agent (tweaking)

12/02/2006: agent runs

12/09/2006: results compiled and all work submitted

Conclusions

As you can see, market-share does make a difference. Thus, the new agent is beating the baseline a majority of the time. As a result, I would stand behind this agent in a competition!

References

[1] Berenson, Mark L., et al. Basic Business Statistics: Concepts and Apllications. Tenth Ed. 679.

[2] Cole, Josh. RationalSCM. Australian National University. .

[3] Collins, John, et al. The Supply Chain Management Game for the 2006 Trading Agent Competition. November 2005. .

[4] Collins, John, et al. MinneTAC. University of Minnesota. .

[5] Collins, John, et al. Component-based Design for a Trading Agent. July 2004. .

................
................

In order to avoid copyright disputes, this page is only a partial summary.

Google Online Preview   Download