Optimizing Online Auction Bidding Strategies with Genetic Programming

1
Optimizing Online Auction Bidding Strategies with
Genetic Programming

• Ekaterina Kate Smorodkina

2
Why Optimize Bidding Strategies?

• Popularity of online auctions
• Limited resources (i.e. )
• Bidding on multiple items increases the
  complexity of the decision making process
• Increasing number of buyers and auction listings
• Difficulty in predicting the behavior of other
  buyers

3
Overview

• Research questions and problem definition
• Online auction overview and auction simulation
• Strategy representation
• Fitness evaluation
• Evolving agents
• Experiments
• Future work

4
Research Questions

• Is it possible to come up with one all-purpose
  bidding strategy for various online auction
  scenarios?
• How successful is genetic programming in evolving
  bidding strategies for online auctions?

5
Online Auctions Overview

• Limited time
• Starting bid
• Email notification
• Identical items for sale
• Unrestricted bidding

6
Online Auction Simulation

• Item listings are randomly created with a
  starting bid and time limit
• Agents are created with random lists of items to
  buy
• Concurrent bidding.
• Retail price on each item is known
• Agents know if they no longer hold the highest
  bid on an item
• Agents are not allowed to go over their account
  balance

7
Strategy Representation

• Expression trees
• Binary operators , -, /, , , max, min
• The size of the trees is controlled by two
  parameters the branching limit and the depth
  limit
• Large trees take longer time to compute a bid
• Input parameters to the expression tree

8
Input parameters to the expression

• Account balance
• Retail price
• Current bid
• Number of items on the list
• Number of items missing
• Sum of the retail prices on the missing items
• The highest bid among all instances of the item
• The lowest bid among all instances of the item

9
Agent Fitness Evaluation

• Maximize the number of items obtained.
• Maximize discount.
• N number of items obtainedM number of items
  on the listR.P retail priceH.B highest bid

10
Evolution Cycle Modified
Initialize
Bid
Evaluate
Select
Compete
Reproduce
Bid
Evaluate
11
Evolving Agents

• Proportional Selection
• Recombination
• Subtree crossover
• Mutation
• Competition
• Elitist strategy
• Termination
• Fitness convergence
• Mutation rate adjustment

12
Experiments to Perform

• Change the environment after each auction round
• Number and characteristics of items in the
  auction
• Agents lists and their initial account balance
• Fitness standard as a way to measure the success
  of the experiment

13
Results

• None yet

14
Future Work

• Finish this project
• Expand the types of operators in the expression
  trees
• Expand input parameters to the expression trees
• Create seller agents