Automatic Trading Agent Design Competition


For my e-commerce class final, I and a team of two others designed and implemented an intelligent program to compete against programs created by other teams. At the end of the semester, the programs (a.k.a. agents) competed against each other to see which could reliably make the highest profit in an product advertisement placement bidding simulation.

We named our agent TigerBlood -- a topical name at the time. By naming our agent for the Charlie Sheen meme, we were hoping for a #winning result. (We did win!)

Our design process was to first implement the simplest possible agent, and then iteratively attempt to beat it in simulation. Our final agent was the 5th version, and was at its core the same idea as the first implementation, as you can read in the final report.

The algorithm essentially hill-climbed the bid in the direction of increased profit -- the bid was either incremented or decremented depending upon whether profit rose or fell from one round to the next.

The TigerBlood agent was very simple compared to research laboratory-submitted agents, but it was able to beat the other agents in the e-commerce class final, and even a couple of the research lab agents! It was clear that in the limited time we had available, the simple and robust solution worked the best.