Multi-Agents System CMSC 691B Gunjan Kalra Peter DSouza

1
Multi-Agents SystemCMSC 691BGunjan KalraPeter
DSouza
2
Outline

• Introduction
• Overview of TAC
• Background
• Strategy used by LARS
• Strategy used by UMBCTAC
• MASTAC strategy
• Design
• Conclusion

3
Introduction

• TAC - trading agent competition held annually
• Strategies used involve single agent per
  competitor
• Winners (Living Agents) used a multi-agent
  strategy that proved to be most effective
• MASTAC Enhancing UMBCTacs strategy by
  incorporating Living Agents architecture

4
Trading Agent Competition

• 12 minute game - 8 competitors per game
• Agents communicate with AuctionBot server
• Each competitor has 8 clients
• 3 commodities - flight tickets, hotel
  reservations and entertainment tickets
• Objective - maximize the total satisfaction of
  the clients
• Learning patterns developed in qualifying rounds

5
Commodities

• Flights
• One flight both ways per day (no in-flights on
  last day, no out-flights on first day)
• Single seller auctions
• Prices set by AuctionBot agent according to
  stochastic function - Perturbations induced in
  price periodically
• Unlimited supply of seats

6
Commodities

• Hotel Reservations
• high priced (Tampa Towers)
• low priced (Shoreline Shanty)
• Ascending auctions
• 16 rooms auctioned off each day
• 16 highest bids accepted
• Cost price for all winners price of 16th
  highest bid Ask price of next set of bids
• Only hotels can sell rooms

7
Commodities

• Entertainment
• Alligator wrestling, amusement park and museum
• Each agent receives allotment of 12 tickets
• 8 tickets per event type allotted
• Agents exchange bids continuously through double
  auctions
• One auction for each event-night combination

8
Scoring

• Penalty of 200 assessed for each ticket owed
  (sell tickets the agent does not own)
• Allocation done by TAC scorer
• Value sum of individual client utilities
• Final score(value of allocation) - (travel
  agents expenses) - (penalty for negative
  entertainment balances)
• Optimal allocation done

9
Background

• General strategies used
• Strategies based on obtaining hotel rooms
• Price prediction algorithm
• Hotel auction models
• Concentrating on individual preferences vs.
  aggregate purchases
• Linear programming solution to find an optimal
  purchase
• Greedy algorithm used in some cases

10
Background

• Risk analysis strategies prior to purchase
• lengthen travel if customer preferences showed
  higher score for attending entertainment
• shorten travel if cost of hotel exceeded penalty
• Analysis of average response time to overcome
  network delay and server performance
• Allocation provided by agents themselves in some
  cases

11
Living Agents Strategy

• 20 agents of 5 different types used
• 1 TACManager
• 5 TACDataGrabber agents
• 8 TACClient agents
• 5 TACAuctioneer agents
• 1 ResultGrabber (offline agent)
• Strategy (never changes) - offer high prices as
  soon as possible, win bids
• No bids withdrawn or changed

12
Living Agents Strategy

• Flight and hotel auctioneers bid for needs of the
  clients only
• Entertainment auctioneers place higher bids for
  buying and lower bids for selling
• Best journey calculated based on
  client-preferences, flight prices and average
  hotel prices

13
UMBCTac strategy

• An adaptive best plan for every client
• Computes TAC value by contrasting every possible
  plan against client preferences
• Larger the TAC value, better the match
• Best plans may change every time new TAC data is
  obtained
• Concentrate on only three clients per cycle due
  to huge computation time

14
UMBCTac strategy

• Sleeps for a minute to compute new bids
• Flights - tickets bought early in the game
• Hotels - offer higher price initially itself
• Entertainment - first find one ticket that had
  maximum entertainment and assign it to a customer
  who is willing to pay for it

15
Design

• Strategy
• one agent works toward obtaining one clients
  best travel plan
• one MASTACmgr agent acts as an interface between
  8 client agents and TAC server
• computes next set of bids based on best bid
  estimates received from each client agent

16
Design

• Flights
• Collect tickets as soon as possible
• Hotels
• Same as that of UMBCTac
• Decision made on per-customer basis
• Entertainment
• Adjust entertainment tickets between agents to
  optimize distribution of tickets

17
Advantages

• Huge search space distributed amongst 8 client
  agents
• Less computation time per agent
• Additional capability to bid for more goods than
  a single agent can
• Better allocation possible due to better travel
  plans for every client

18
Conclusion

• Performance benefits observed in real-life
  scenarios with longer time duration
• Implementation results still to be obtained
• Less overhead in communication time as compared
  to Living Agents strategy
• Higher throughput as compared to UMBCTac strategy