TAGA An Advanced Trading Agent Framework Youyong Zou, Li Ding, Harry Chen, Rong Pan, Tim Finin University of Maryland, Baltimore County April 2003

Overview The Trading Agent Competition has advanced the theory and practice of agent-based dynamic markets TAC has an abstract and overly simplified model of multiagent systems. TAGA is an enhanced environment for experimenting with agent-based dynamic markets using important new technologies: FIPA standards for agent communication OWL for ontologies and content DAML-S for describing and reasoning about services Taga is intended as a research and teaching testbed for both dynamic markets and agents and the semantic web Taga was recognized as the best student entry in the 2003 EU agent technology competition.

Outline Review of TAC The need for a more realistic model The TAGA model Semantic web in TAGA Future directions Conclusions

Review of TAC

TAC The Trading Agent Competition proposed (1999) and first run (2000) by Mike Wellman and Peter Wurman International competitions in 2000, 2001 and 2002 were based on a simple travel scenario 2003 will have two competitions: TAC classic – travel agency scenario TAC SCM – supply chain management scenario The finals will be held at IJCAI in August See http://www.sics.se/tac/

TAC Classic In a game instance, 8 travel agencies compete to satisfy their clients, buying flights, hotel rooms and entertainment tickets in various markets.

TAC Classic Trading Agents TAC Server Market server Game Spectators Communication with trading agents Market server Information database Publish data via web and applet TAC Server Trading Agents Game Spectators Living Agents ATTac 006 TAC Classic

Market Game Rules: Packages Inbound flight Days 1-4 Outbound flight Days 2-5 Hotel reservation Days 1-4, Good or Cheap Entertainment tickets Days 1-4, AP, AW, or MU 4 + 4 + 4*2 + 4*3 = 28 auctions

Client Utilities Utility=1000-travelPenalty+hotelBonus+funBonus IAD: Ideal Arrival Date IAD=1..4 IDD: Ideal Departure Date IDD=2..5 HV: Value for Good Hotel HV=50..150 AP: Amusement Park AP=0..200 AW: Alligator Wrestling AW=0..200 MU: Museum MU=0..200 Utility=1000-travelPenalty+hotelBonus+funBonus Travel Penalty = 100*(|IAD-AD|+|IDD-DD|) hotelBonus = HV, if Good hotel funBonus = AP + AW + MU, if package includes resp. entertainment

Auctions: Flights Unlimited supply, continuous clearing Initial price within $[250;400], max = $800 Every 30-40 sec, price perturbed by a random value drawn from [-10; x(t)] Perturbations biased upwards: x(t) = 10+(C – 10)*(t / 720), C is a number, drawn uniformly from [10;90]

Auctions: Hotels Seller makes available 16 rooms per night Rooms sold in ascending, multi-unit, 16th price auctions: Top 16 bids win winners pay 16th highest price Hotel auctions close randomly every minute to avoid late bidding

Auctions: Entertainment Continuous double auctions Each agent receives random initial endowment of tickets

The need for a more realistic model

Real World Issues TAC classic assumes that agents interact via a few centralized markets. Technology is basically client-server with well defined APIs and simple XML encodings. Real word interactions are varied and rich Customers can chose to interact with travel agencies, dynamic markets, or directly with service providers Choices are governed by value, speed, reputation Bilateral negotiation for information and services is common Rich information exchange abounds -- customers have complex interests and preferences, service providers have detailed descriptions, etc. Common ontologies are important Trust and reputations are important.

Our Research Objective To develop an agent framework for building simulations of the traveling business Using the Agentcities to support the Trading Agent Competition in a global scale Using the Semantic Web languages & tools to enable ontology sharing and reasoning Extending the FIPA protocols to support open market auctions and agent negotiations

TAGA

Travel Agent Game in Agentcities Game Objective: develop strategies for different agents to achieve their objectives You can choose to implement: Customer Agent Web Service Agent Travel Agent OR OR

The UMBC TAGA Demo What we have developed and achieved Travel Agent Game in Agentcities (TAGA) A FIPA compliant agent framework that extends and enhances the Trading Agent Competition (TAC) Project won the Best Student Entry in the Agentcities sponsored Agent Technology Competition held in Feb. 2003 in Barcelona Our main contributions Auction services are developed to enrich the Agentcities environment The use of Semantic Web languages (OWL) improves agent interoperability DAML-S ontology is employed to support service registration, discovery and invocation. The development of TAGA is an attempt to bring Trading Agent Competition (TAC) into a more realistic e-commerce world. In TAGA, agents can use auctions and other markets but can also interact directly with one another, buying, selling, querying, etc. Our objectives include: (1) Developing an open framework for building trading game simulation; (2) Developing a MAS research testbed for studying complex intelligent agent behavior; and (3) Developing a MAS teaching tool that helps beginners to have a hands-on experience in building software agents.

The TAGA Game and Players Game Objective: develop strategies for different agents to achieve their objectives Human players can choose to implement/play agents with different objectives: Customer Agent (CA) find travel arrangements try to save $$ Travel Agent (TA) satisfy customers’ needs maintain a good reputation try to maximize profits Web Service Agent (WSA) sell “goods” (e.g., plane tickets, hotel rooms, entertainment tickets) TAGA provides default agent shells Anyone can download a copy of the default agent, customize it and participate in the game. Users can also develop customized agent strategies to compete with other agents. Agents must follow well defined, industry supported standards for agent communication languages and protocols Based on TAGA, researchers can develop more advanced agent models involving distributed trust, social norms, reputation models, security and/or privacy considerations.

TAGA in Action Our Demo Will Show … TAGE Home Page http://taga.umbc.edu Create a TAGA game online Download the latest TAGA pkg and docs http://taga.umbc.edu/taga/download/ TAGA on Agentcities network (UMBCTac.agentcities.net) Baltimore, MD USA http://www.agentcities.net/ Our Demo Will Show … The architectural design of TAGA including ontologies, agent communication languages, negotiation protocols and agent development shells. The runtime features of a TAGA game server -- a multi-tier web application involving web severs, remote Java applications, FIPA platforms, and the semantic web. The runtime interactions of different TAGA agents including the complete run of a typical TAGA scenario. How to customize and extend the TAGA framework for research and teaching purposes TAGA supports heterogeneous agent platform. A FIPA-JADE agent can interact with a FIPA-AAP agent

A Typical Scenario Hotel WS Airline WS Market Oversight Agent Bulletin 4 Airline WS 6 Market Oversight Agent 3 Bulletin Board Auction Service (1) Every 30 sec., a customer agent (CA) with random travel profiles posts a request on the Bulletin Board. (2a) The Bulletin Board forwards the request to all registered TA (calling for travel package proposal) (2b) CA considers incoming proposals submitted by different TA using its own utility function & chooses 1 winning TA. (3) The chosen TA attempts to put together a travel package for the CA using its built-in strategies (e.g., buying plane tickets through auction & reserving hotel rooms directly from the online web services) (4) After each successful transaction, $$ are deposited to the accounts of the WS in the Market Oversight Agent (5) Once TA has made a complete travel package, it informs its client CA. (6) Upon receiving the travel package, CA deposits $$ into the TA’s account in the Market Oversight Agent. CA TA 1 2a 2b 5

The TAGA Game Server http://taga.umbc.edu/taga/play/demo.htm View TAGA game status View ACL message traffic Monitor Open Market Auction TAGA is constantly running as part of the Agentcities network (http://www.agentcities.net/), which connects 134 registered agent platforms and agent services The TAGA implementation is robust and flexible. TAGA agents can dynamically join and leave the system without interrupting the execution of other agents in the system. TAGA has a robust Apache + MySQL web back-end system and a highly customizable PHP web scripting front-end. Agents that are implemented in either JADE and AAP can interact with the server. Monitor Customer Agents (and more … new agent, user login, create game, game history)

TAGA Now and in the Future Add more reasoning to TAGA agents Use OWL-S for service discovery and planning Add more demonstration agents Develop toolkit for teaching agent and SW technology Open Source Now Built on FIPA standards: Agentcities + April Agent Platform (AAP) + JADE Employs OWL ontologies in agent communication Uses SOAP & WSDL in Web Service registrations Robust & persistent web server backend (MySQL + PHP + Apache) We have successfully demonstrated TAGA and won a prize in the Agent Technology Competition in Feb. 2003, which was held in conjunction with the third Information Day organized by the Agentcities.NET project. For more information see http://taga.umbc.edu/ for Mailing lists News & documentation Contact info and more … For more info: http://taga.umbc.edu/

Prototype Implementation Travel Agent Contracts Static contract: the system assigns each TA with a fixed number of CA (in TAC-02) Dynamic contract: Each TA bids for its own clients through the Bulletin Board Agent Auction Service Agent English & Dutch auction “Name your price” auction (priceline.com & hotwire.com)

TAGA and the Semantic web

TAGA and the semantic web Semantic web languages and tools are used in several ways in TAGA To specify and publish the underlying common ontologies As a content language within the FIPA ACL messages As the basis for agent KBs via an XSB-based reasoning tools To describe and reason about services

TAGA Ontologies Travel.owl – travel related concepts Auction.owl – models for different kinds of auctions and markets Fiapowl.owl – Owl as a FIPA compliant content language Tagaql.owl – a simple query language

Examples…

Future work and conclusions

Future Work Package TAGA for Open Source AAP/JADE agent shells for various kinds of agents: CA, TA & WSA TAGA protocols for registration, logging etc. Use OWL-S for service description and matching Provide new classes of agents, e.g. a reputation server Exploit ontology reasoning

TAGA Significant Features An extensible and open framework for the Trading Agent Competition (TAC) TAGA explores the use of Semantic Web languages (OWL) in a multi-agent environment. Persistent gaming service Players can dynamically join & leave the game UMBC TAGA platform has been running on the Agentcities network for more than 100 days

For more information http://taga.umbc.edu/