Technion Israel Institute of Technology Department of

Technion – Israel Institute of Technology Department of Electrical Engineering המעבדה לבקרה סמסטר חורף תשס"ב הצגת פרוייקט Autonomous Bidding Agent in the Trading Agent Competition. יבגני טלביצקי, אהרון זלצבורג מנחה: שי מנור 1

Talk Outline TAC Introduction to Game strategy Agent behavior parameters Algorithm description A bidding example The conclusions 2

TAC History: The first trading agent competition (TAC), held in Boston, Massachusetts, on 8 July 2000. TAC was organized by a group of researchers and developers led by Michael Wellman of the University of Michigan and Peter Wurman of North Carolina State University. The participants were challenged to design a trading agent capable of bidding in online simultaneous auctions for complimentary and substitutable goods. 3

Why Autonomous Agents ? • Improved decision processes. • Better quality and more coherent decision-making. • Reduced training and supervisory time. 4

General Game Description • A TAC game instance lasts 12 minutes and pits eight autonomous bidding agents against one another. • Each TAC agent is a simulated travel agent with eight clients, each of whom would like to travel from TACtown to Tampa and home again during a five-day period. • Each client is characterized by a random set of preferences for arrival and departure dates, hotel rooms, and entertainment tickets. • A TAC agent’s objective is to maximize the total utility (a monetary measure of the value of goods to clients) minus total expenses. 5

Project Goals • Create a program that will participate in TAC competition. • Implement a strategy that determines how the agent makes his decisions. • Find the key parameters that influence the agent’s behavior. • Define the behavior parameters in the way that will provide the best performance. • Make an agent that will be competitive with other trading agents participating in the TAC. 6

Introduction to Game strategy Flight Auctions Hotel Auctions Entertainment Auctions 7

Client Utility Functions • u = 1000 - travel_penalty + hotel_bonus + + fun_bonus • A client receives zero utility for an infeasible package. • The range of possible utilities for a feasible package assigned to a particular client. Arr Dep Hotel AW AP Umin = 1000 - 300 + 0 Umax = 1000 - MU + 0 + 0 = 400 0 + 150 + 200 = 1750 8

Hotel Auctions • Intensive bidding start time point. • Bidding for two hotel types. • Give up if its not worthy to buy. • Increase price according to the amount of hypothetically lost bids. 9

Flight Auctions • Bid at the beginning. • Prefer a chipper ticket to client preferences. 10

Entertainment Auctions • Half bonus price. • Sell price restrictions. • Sell one a time. 11

Agent behavior parameters • Nervousness. • Aggressiveness. • Uncertainty. • Greediness. • Risk. 12

Algorithm description 1. Connection to server 2. Get game parameters and clients initial distribution 3. Initialize primary target package 4. Main loop a. Check transactions and Assign bids b. Change targets c. Update bids d. Calculate orders e. Buy tickets: Ø Buy flights Ø Buy hotels Ø Buy entertainments f. Sell entertainment g. Recalculate agent parameters 13

A bidding example 14

15

Conclusions • Parameters calibration • Server latency dependence • Project achievements Thanks 16

17