Download presentation
Presentation is loading. Please wait.
Published byElijah Morton Modified over 9 years ago
1
David Pardoe Doran Chakraborty Peter Stone The University of Texas at Austin Department of Computer Science TacTex-09: A Champion Bidding Agent for Ad Auctions
2
Ad Auctions
5
Which keywords to bid on? –Who is searching for what? –Who am I advertising against? How much to bid? –What are others bidding? –What position will I get? –How many clicks and conversions will I get? What ads to display? How to monitor my advertising campaign? –What feedback is available? –Use spending limits?
6
Background Much work on mechanism design problem –Varian 2007, Edelman et al. 2007 Work from an advertiser’s perspective focuses on isolated subproblems (often stylized) –keyword selection: Rusmevichientong and Williamson 2006 –multi-auction bidding: Zhou and Naroditskiy 2008 –predicting clicks: Richardson et al. 2007 Trading Agent Competition – Ad Auctions –solve full bidding problem against other researchers –designed by U. Michigan in 2009 –follows other successful TAC competitions
7
Outline Introduction TAC/AA Overview The TacTex Agent for TAC/AA Competition Results Experimental Results Conclusion
8
Ad Auction Agents AdvertiserPublisherUser Competition EntrantsEnvironment (Built-in)
9
Products and Queries 9 products Query format: (manufacturer, type) –either may be null –16 total queries
10
User Behavior Each user interested in one product Users cycle through states –not searching, 4 levels of searching –increasing query specificity, chance of buying Searching users submit one query daily –user sees up to five ads (impressions) –may click an ad (more likely at higher positions) –may make a purchase (conversion)
11
Game Format 8 advertiser agents per game 60 game days, 10s each Each day, for each of the 16 queries, advertisers: –submit a bid (per click), spending limit, and ad –receive own outcomes: impressions, clicks, conversions, costs –see limited information on other advertisers: average position when ad was shown Agents have limited capacity, product specialties
13
Outline Introduction TAC/AA Overview The TacTex Agent for TAC/AA Competition Results Experimental Results Conclusion
14
TacTex Agent Overview
16
User Model Particle filters for each product (users per state) Filtering based on daily impressions Update based on known user transition dynamics Likelihood = Probability of observed impressions (binomial distribution) Likelihood = Probability of observed impressions (binomial distribution) d – 1 impressions d – 1 impressions Particle Filter for a particular product Particle updated user population updated user population
17
- users in one state for one product type
19
Advertiser Model Estimate bids of other advertisers Average of two estimators First estimator: –particle filter for each query –joint distribution over all advertiser bids Second estimator: –distribution over discrete bids –separate distribution for each query, advertiser –model probability of bid transitions Also estimate spending limits
21
Two-level Optimization Goal: determine bid and spending limit for each query to maximize future profit Predicted bids and impressions for each query Capacity, desired conversions Greedy Optimizer Optimal bids and resulting profit Single Day Optimizer: Multi-Day Optimizer: Hill climbing search Single Day Optimizers Proposed conversion goal for each remaining game day Expected profit
22
Outline Introduction TAC/AA Overview The TacTex Agent for TAC/AA Competition Results Experimental Results Conclusion
23
Competition Results IJCAI 2009 15 teams Final round: top 8 agents, 80 games 1. TacTex 79,886 2. AstonTac (Aston U) 76,281 3. Schlemazl (Brown U) 75,408 4. QuakTac (U Pennsylvania) 74,462 5. Munsey (U Washington Tacoma) 71,777 6. epflAgent (EPF Lausanne) 71,693 8. UMTac (U Macau) 66,930 7. MetroClick (CUNY & Microsoft) 70,632
24
Competition Results AstonTAC and Schlemazl: –slightly higher revenue per conversion –much higher cost per click Other agents: –lower cost per click –much lower revenue per conversion TacTex struck right balance
25
Outline Introduction TAC/AA Overview The TacTex Agent for TAC/AA Competition Results Experimental Results Conclusion
26
Experiments 7 other agents from Agent Repository One (modified) TacTex 50 games per experiment Most important (> 3000 drop in score): –no multi-day optimization –not estimating spending limits Moderately important (> 400 drop in score) –add noise to estimated bids of others –add noise to estimated spending limits of others –add noise to own bids (single day optimizer) –no user model
27
Conclusion and Future Work TacTex a complete agent for ad auctions Estimates/predicts all values of interest Optimizes with respect to these values All agent components contribute to performance Future work: improve advertiser modeling –machine learning to improve bid estimation –predict future bids given estimates
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.