Designing a Multi-Agent Portfolio Management System

Designing a Multi-Agent Portfolio Management System

Keith Decker, Katia Sycara, and Dajun Zeng

The Robotics Institute, Carnegie Mellon University

5000 Forbes Ave, Pittsburgh, PA 15213

(decker,sycara,zeng+)@cs.cmu.edu

Abstract

The voluminous and readily available information on the Internet

has given rise to exploration of Intelligent Agent technology

for accessing, filtering, evaluating and integrating information.

In contrast to most current research that has investigated

single-agent approaches, we are developing a collection

of multiple agents that team up on demand, depending

on the user, the task and the situation, to access, filter and integrate

information in support of user tasks. We are investigating

techniques for developing distributed adaptive collections

of information agents that coordinate to retrieve, filter and fuse

information relevant to the user, task and situation, as well as

anticipate user’s information needs. Our approach is based on

(1) case-based task and situation models, (2) flexibleorganizational

structuring, and (3) reusable agent architecture. Weare

currently implementing the system in the domain of financial

portfolio management. While this paper focuses on the big

picture, the final section will describe the current implementation

and point to our work on detailed technical solutions.

 

Introduction

Due to advances in technology, diverse and voluminous information

is becoming available to decision makers. This

presents the potential for improved decision support, but

poses challenges in terms of building tools to support users

in accessing, filtering, evaluating and fusing information from

heterogeneous information sources(Kuokka & Harada 1995;

Arens et al. 1993; Collet, Huhns, & Shen 1991). Most reported

research on Intelligent Information Agents to date has

dealt with a user interacting with a single agent that has general

knowledge and is capable of performing a variety of user

delegated information finding tasks (e.g., (Etzioni & Weld

1994; Maes 1994)). For each information query, the agent is

responsible for accessing different information sources and integrating

the results. We believe that, given the current computational

state of the art, a centralized agent approach has

many limitations: (1) a single general agent would need an

enormous amount of knowledge to be able to deal effectively

with user information requests that cover a variety of tasks, (2)

a centralized information agent constitutes a processing bottleneck

and a “single point of failure”, (3) unless the agent has

beyond the state of the art learning capabilities, it would need

considerable reprogramming to deal with the appearance of

new agents and information sources in the environment, (4)

because of the complexity of the information finding and filtering

task, and the large amount of information, the required

processing would overwhelm a single agent. Because of these

reasons and because of the characteristics of the Internet environment,

we employ a distributed collaborative collection of

agents for information gathering.

We are currently working on a system where each user is

associated with a set of agents which have access to the task

and situation models and keep track of the current state of

the task, situation, environment and user information needs.

Based on this knowledge, the agents decide what information

is needed and initiate collaborative searches with other agents

to get the information. During search, the agents communicate

with each other to request or provide information, find

information sources, filter or integrate information, and negotiate

to resolve conflicts in information and task models. The

returned information is communicated to display agent or

agents that possibly combine it with information from other

sources (e.g. the user) and/or filter it for appropriate display

to the user.

This paper focuses on the design of such a system of agents

for the task environment of financial portfolio management,

and on the key issues that we will be addressing. These issues

include: (1) gathering and integrating diverse information

sources with collaborating software agents, (2) case-based

user, task and situation models, (3) adaptive integration of

planning, coordination, scheduling, and execution. In our

system, case-based reasoning provides meta-level control and

activation of agents. Depending on the task, user and situation,

case based retrieval selects current planning goals, information

needs and information gathering goals. Based on the

plans and information gathering goals, agent teams are activated

“on demand” to access, integrate and filter information

to fulfill these goals. New information can be incorporated

in the case base and may give rise to new plans and information

gathering goals (and as a result activation of potentially

different agent teams). The system has three types of

agents, task agents, information agents, and interface agents.

Task agents have information about tasks and associated information

gathering goals. Information agents havemodels of

information sources, information access strategies and associated

task agents to whom the information should be returned.

Interface agents interact with the user receiving user specifications

and delivering results. They acquire, model, and utilize

user preferences to guide system coordination in support of

the user’s tasks. This work is a continuation of our previous

work on multi agent information access, filtering and integration

of everyday organizational tasks (Sycara & Zeng 1995).

Of the three types of agents, information agents are the

most well-defined. Each information agent—by definition—

shares a set of reusable behaviors that give it the ability to

process one-shot and periodic queries, monitor for certain

conditions, advertise it’s capabilities appropriately, and react

reasonably in the event of a local or external data source failure.

If a domain content ontology has been fixed beforehand,

an information agent can be created quickly with little

programming—it requires a database definition, and some

code to access the external data source (i.e. web page, newsgroup,

or normal database) and create local database records.

Information agents are described in more detail in (Decker,

Williamson, & Sycara 1996).

The Portfolio Management Domain

To evaluate our domain independent agent control, organization,

coordination and architectural schemes, we have chosen

financial portfolio management as a task domain. This is the

task of providing an integrated financial picture for managing

an investment portfolio over time, using the information

resources already available over the Internet. This task environment

has many interesting features, including: (1) the

enormous amount of continually changing, and generally unorganized,

information available, (2) the variety of kinds of

information that can and should be brought to bear on the

task (market data, financial report data, technical models, analysts’

reports, breaking news, etc.), (3) the many sources of

uncertainty and dynamic change in the environment.

The overall task in the portfolio management domain, as

stated by modern portfolio theory (Markowitz 1991), is to

provide the best possible rate of return for a specified level of

risk, or conversely, to achieve a specified rate of return with

the lowest possible risk Risk tolerance is one of the features

that characterize the user of our system; other features include

the user’s investment goals (long-term retirement savings?

saving for a house?) and the user’s tax situation. In

current practice, portfolio management is carried out by investment

houses that employ teams of specialists for finding,

filtering and evaluating relevant information. Current practice

as well as software engineering considerations motivate

our multi agent system architecture.

Previous work in the portfolio management domain (see

(Trippi&Turban 1990) for one collection) has focused on the

portfolio selection process (i.e., “stock picking”) as opposed to

portfolio monitoring—the ongoing, continuous, daily provision

of an up-to-date financial picture of an existing portfolio.

A multi-agent system approach is natural for portfolio

monitoring because the multiple threads of control are a natural

match for the distributed and ever-changing nature of

the underlying sources of data and news that affect higherlevel

decision-making processes. A multi-agent system can

more easily manage the detection and response to important

time-critical information that could appear suddenly at any

of a large number of different information sources. Finally,

a multi-agent system provides a natural mapping of multiple

types of expertise to be brought to bear during any portfolio

management decision-making. Existing DAI techniques for

resolving conflicting opinions, negotiation, and argumentation

can be brought to bear on these problems (Sycara 1989).

The overall portfolio management task has several component

tasks. These include eliciting (or learning) user profile

information, collecting information on the user’s initial portfolio

position, and suggesting and monitoring a re-allocation

to meet the user’s current profile and goals. As time passes, assets

in the portfolio will no longer meet the user’s needs (and

these needs may also be changing as well). Our initial system

focuses on this ongoing portfolio monitoring process.

The task of monitoring individual portfolio assets gives rise

to a variety of concurrent goals, such as monitoring an asset

currently being held (should we continue to hold it? sell some

or all of it?), or monitoring the buying or selling of an asset.

Buying and selling at this high level are not instantaneous

transactions, but also require careful planning and monitoring

of plan execution. For example, let us assume that in the

process of monitoring the system comes to recommend that

an asset be sold. The way in which it is sold will be determined

in part by the reason for the sale—perhaps the asset

is no longer performing as required for it’s role in the portfolios

asset allocation mix. Perhaps it is performing too well,

and there is a growing possibility that the asset has reached a

peak. In this second case it is often prudent not to sell one’s

entire holdings all at once, but to sell in phases and place the

appropriate standing orders to protect the user from a sudden

downturn (while avoiding worry over simple daily price

fluctuations) Thus the goal of selling an asset is not one that

requires only a simple short sequence of actions, but one that

requires careful planning and monitoring of that plan as it is

executed, over an extended period of time.

Case Based Situation and Task Models

We are proposing to model the information gathering task

as a planning task where planning and execution are interleaved

and search is guided by user, task and situation models.

We believe that cases that incorporate user, task and situation

models can effectively provide instantiations of the situationdependent

task structure and the associated team of information

gathering agents. The case base contains cases of successful

and unsuccessful information gathering episodes and

information evaluation. After each information gathering cycle

the case base is updated. Thus, learning is integrated

with problem solving and is achieved automatically. This

feature makes Case-Based reasoning very preferable in application

domains with open and dynamically changing world

model (Sycara, Zeng, & Miyashita 1995). Case based reasoning

offers three general advantages. First, since it relies

on reusing specific experiences rather than reasoning from a

general world model, it provides for more efficient planning.

Second, since it can start with few cases in memory and incrementally

acquire new cases based on a reasoner’s interactions

with the world, it makes few assumptions about the completeness

and correctness of world knowledge. Third, previous past

failures warn the planner about the possibility of failure in

current circumstances and, hence help avoid future failures.

Each case will be automatically labeled with identifying information

such as user’s name (a user could be a software

agent), time of creation, time of modification, etc. So, a case

will carry a complete audit trail of its origin and modifications.

This information is essential for careful analysis of all

the planning knowledge that is exchanged during system operation,

and we will rely heavily on it during the initial stages

of knowledge acquisition and development. In addition, this

information can be used to evaluate the performance of the

system under different planning scenarios, since each item will

be clearly identified.

The library of cases we propose can be viewed as a casebased

scheme for meta control that offers agents access to task

and situation specific information. The agents use this information

to focus search and filtering. The asynchronous

information gathering activity of the agents results in new information

about the world (new cases) that gets incorporated

into the case base. Case base updates result in formation of

new memory indices. These new indices, along with any new

user inputs (e.g., information gathering requests, change in

context) activate a new set of cases that reflect possibly new information

seeking goals, and new sets of agents. In this sense,

the case base can be viewed as (1) tracking the user’s intentions

and his/her evolving information needs, (2) reflecting changing

situations, (3) recognizing new events, and (4) learning

information retrieval tactics/heuristics.

Figure 1 shows an active case base which receives as input

notification of events, either directly from the user or

from new information that becomes available (e.g., fromother

software agents). Updates of the case-base with the addition

of new information finding episodes and new indices

are also considered an event that results in the activation of

a new round of reasoning. In the following, we use an example

to illustrate the general Case-based agent invocation process.

Given a situation, the case-base calls a certain task assistant,

say Analyst Tracking Agent, which the reasoning process

determines is relevant to meeting the information gathering

goals (which in turn might be information requests from Portfolio

Manager Agent). Portfolio Manager Agent calls upon

Earnings analysis Agent, News Classifier, etc., to locate information

from the infosphere directly (in the case of News Classifier)

or indirectly (in the case of Earning Analysis Agent).

Note that there is no hierarchy of agents here: in a different

situation Earning Analysis Agent might have been called directly

by the case-base reasoning process. After collaborating

to find and filter information, Analyst Tracking Agent updates

the case base with new information finding episodes that include

a timestamp and the results of the search. If unexpected

news (“unexpected” in terms of the current situation) has been

found during information retrieval (such as major corporate

merge news), the case-based process will interrupt the regular

plan execution and take other emergency actions. For example,

the case-base process might invoke the interface agent to

notify the user right away.

From the perspectives of Case-Based reasoning, there are a

number of important research questions that we need to address.

The most fundamental question is what constitutes a

case? A case in case memory describes a specific information

scenario including: (1) information needs and goals which

might come directly from the user or from other software

agents, (2) global features which give an abstract characterization

of the situation in which this information gathering

operation takes place, (3) features of local nature which describe

in detail the information about information sources,

inter-agent interactions, etc., (4) information retrieval plan

skeleton, (5) feedbacks/evaluations with respect to acquired

information and information retrieval effectiveness/efficiency,

and (6) potential failures and re-trial, planmodification information.

At the most abstract level, we might index an asset

monitoring task plan fragment using several broad situation

characteristics such as asset type, industrial sector, ownership

records, tax status, user profile, etc. These characteristics are

represented in the case base and used as indices for the retrieval

of plan fragments, associated information gathering

goals and associated lists of agents to be activated. We are

currently investigating other important issues such as efficient

case indexing mechanism, case retrieval/matching approaches,

and initial case collection.

Organizational Structure

We propose a general system organization in which agents are

directly activated based on the top-down elaboration of the

current situation. These agent activations, guided by casebased

retrieval according to the current situation, dynamically

form an organizational structure that fits in with the user’s

current profile, tasks, and other situational features. This organization

will change over time, but will also remain relatively

static for extended periods (for example, while monitoring

currently held investments during stable market periods).

Information that is important for decision-making (and

thus might cause an eventual change in organizational structuring)

is monitored at the lowest levels of the organization

and passed upward when necessary.

In this type of organization (see Figure 1), “task agents”

or “task assistants”(Sycara & Zeng 1995) continually interleave

planning, scheduling, coordination, and the execution

of domain-level problem-solving actions. Task agents interact

with one another and with “information agents” or “information

assistants” that encapsulate network information sources.

Task agents retrieve, coordinate, and schedule plans based on

local knowledge modulated by situational context. A task assistant

decomposes an information request into information

seeking goals and subgoals and interacts with the information

assistants to gather the information. In this architecture, a

task assistant does the final filtering and fusing of information

before it passes it on to agents above it in the organizational

structure (requesting agents). This incremental information

fusion and conflict resolution increases efficiency

and potential scalability (e.g., inconsistencies detected at the

information-assistant level may be resolved at that level and

not propagated to the task-assistant level) and robustness (e.g.,

whatever inconsistencies were not detected during information

assistant interaction can be detected at the task-assistant

level). In addition, a task assistant composes a new case that

incorporates its findings to be stored in the case memory.

In this architecture, information-assistants would have

models of their associated information sources, learn the reliability

of those sources, as well as strategies for low-level information

fusion and multiple methods for responding to information

requests. As an example of the latter, a stock ticker

monitoring agent might have several methods available to it

that trade off time, cost, and quality:

_ one or more sources of 15-minutedelayed values (with

varying reliabilities and average response delays)

_ one or more sources of real-timequotations that charge a

fee (more reliable response but still not guaranteed)

_ the ability to guess a quotebased on recent data and simple

models (very fast but of low quality).

On the other hand, task-specific assistants have a model of

the task domain, executable methods for performing the task,

knowledge of an initial set of information-assistants relevant

to their task and strategies for learning models of pertinent

information-assistants.

Figure 1 shows a top-level portfolio manager agent which

receives as input notification of events, either directly fromthe

user or from the case base, or from information that becomes

available (e.g., from task and information agents). Given the

current situation, the portfolio manager agent:

_ instantiates task plans andassociated information gathering

goals according to the current situation

_ coordinates those plans with otheragents (this includes task

assignment actions that activate task assistants)

_ schedules and monitors theexecution of its local actions.

In Figure 1, the fundamental, technical, news, and outsideanalyst

task agents have been activated in this manner. These

agents are task assistants that can either locate information via

information assistants, or by calling upon other task agents.

There is not a strict hierarchy of agents—the same task and

information agents may be called upon by different parts of

the portfolio management organization. After collaborating

to find and filter information, task agents update the case base

with new information finding episodes that include a time

stamp and the results of the search.

This architecture has potential advantages and drawbacks.

The advantages include:

_ There is a finite number of taskassistants that each agent

communicates with.

_ Because information processing isdone by all the task

agents at every level (rather than by having one task agent

receive all data fromevery information agent) we avoid having

a single computational bottleneck point.

_ The task assistants areresponsible for checking information

quality, filtering irrelevant information, recognizing important

information, and integrating information fromheterogeneous

information sources for their respective tasks.

_ The task assistants areresponsible for activating relevant

information assistants and coordinating the information

finding and filtering activity for their task.

All the above characteristics, by imposing some structure

through definition of task assistants, contribute to overall system

responsiveness. On the other hand, there are potential

drawbacks:

_ The portfolio manager is a singlepoint of failure. Such failures

can be mitigated by expending the resources to have,

for example, a redundant portfolio manager that takes over

in case of failure.

_ Each task assistant alsoconstitutes a “single point of failure

for that task”. This can be mitigated by having more than

one task assistant (either clones of each other or not). In the

case where two different assistants exist for the same task,

the task assistants must negotiate to resolve inconsistencies.

We propose to explore the use of negotiation strategies for

resolving inconsistencies.

Analyst

Tracking Agent

Breaking

News Agent

Technical

Analysis Agent

Fundemental

Analysis Agent

Portfolio

Manager

Agent

Earnings

Analysis Agent

Ticker Tracker

Market Tracker

News Classifier

Historical

Stock Prices

Historical market

Information

Economic

Indicator Tracker

Infosphere

Task Agents Information Agents

Active Case Base

Agent Activation

Figure 1: A “direct invocation” agent organization for a portfoliomanagement system.

In the stock portfolio example, task assistants for areas such

as earnings analysismight be replicated and allowed to specialize

on various industry groups (one agent to handle banking

industry earnings, one for manufacturing, etc.). Such agents

might begin as clones, but learn specialized case information.

In the event of a failure, a non specialist would still be able to

retrieve useful plans for a task inside its area of expertise, but

outside of its specialty.

The Portfolio Monitoring Task

We can represent the plans that are retrieved using TÆMS task

structures (Decker 1995). TÆMS task structures are based on

abstraction hierarchies, where task plans are elaborated via a

“subtask” relationship into acyclic directed graphs that have

actions, called executable methods, at their leaves. Such structures

are compatible with most planning representations, and

provide the necessary information both for scheduling activities

that arise from multiple plans, and for coordinating the

activities of multiple agents. We have in fact constructed

a decision-theoretic hierarchical task network planner using

extensions of this representation framework (Williamson,

Decker, & Sycara 1996). The extensions include the ability

to represent and reason about periodic tasks. As shown

in Figure 2, a top level portfolio management agent interacts

graphically and textually with the user to acquire information

about the user’s profile and goals; as mentioned earlier, we

will assume in this paper that the system has gone through an

initial usage period and has reached a “steady state” of monitoring

the current portfolio.

Such a monitoring task includes gathering opinions from

various task experts, integrating this information, and then

making or updating the recommendation (such as buy, sell,

or hold) for the asset under consideration. These tasks are

persistent, in that they are continuously present. An agent

will be dealing with many such tasks simultaneously. Gathering

opinions from the area experts (fundamental analysis,

technical analysis, news, and the opinions of other analysts—

the published output of similar human organizations) requires

registering with them and then either waiting for new opinions

to be received or asking for them directly. An opinion

consists of not just a buy/sell/hold recommendation but a

short list of positive and negative reasons for holding that

opinion, and potentially both symbolic and numeric measures

of uncertainty. Information integration involves removing

redundant information, resolving conflicts (or declaring

them unresolvable), and forming a coherent group opinion,

that can then be used for decision-making (in the light of the

user’s risk tolerance, investment goals, asset allocation, tax status,

and so on). The conflict resolution process may involve

negotiation between the agents involved.

monitor

and hold

stock

gather

opinions

integrate

information

make

recommendation

remove

redundancy

T

min

abstract task with quality

accrual function min

subtask relationship

enables relationship

facilitates relationship

Fund. Op.

Update

Provide

fundemental

opinion

update

gather

opinions

integrate

information

make

recommendation

— Management quality

— P/E Ratios

— Earnings Reports

S

Provide

immediate

opinion

max

(OR)

Portfolio Manager Agent

Fundemental

Analysis

Agent

Technical

Analysis

Agent

Breaking

News Agent

Analyst

Tracking

Agent

Other

Agents

S

resolve

conflicts

form

coherent

opinion

assess user

profile

information

assess

overall

goals

assess tax

situation

Figure 2: A task structure representing high-level portions of themonitor-stock-and-hold task.

An Example of Coordination: Earnings Report

Interpretation

One interesting subproblem in portfolio management is acquisition

and interpretation of earnings reports and earnings

estimates. The earnings analysis task is a complex one that includes

estimating the impact of one company’s earnings on

other companies in a sector, the information contained in

one company’s earning report that actually releases information

about all companies in a sector, timing in the release of

earnings reports (especially for smaller companies), and the

differences in actual earnings versus expectations. It is important

to track revisions in earning estimates over time, as they

often give important clues as to future price moves.

Figure 3 shows a relationship between the abstract plans

of the earnings analysis agent and the human analyst opinion

tracker agent. The earnings analysis agent initially needs to

get data on a companies current and historical earnings patterns,

and then it needs to keep up to date on new earnings

reports as they are released. Not only does it need to track the

new earnings of the company in question, but also the earnings

of other companies in an industry sector. For example, a

change in the portion of earnings attributed to sales is often

applicable to all companies in a group, unlike changes to costs

(sales minus earnings) (Joh & Lee 1992).

The analyst tracking agent gathers, from news and other

sources, existing and updated analyst reports on a company,

including revised earnings estimates (often part of a larger

report). This part of the data, if transmitted to the earnings

analysis agent, can somewhat speed up (i.e., facilitate)

the process of gathering earnings expectations. We have

demonstrated the use of general coordination mechanisms,

called the GPGP (Generalized Partial Global Planning) approach,

that can easily coordinate such task structure interactions

(Decker & Lesser 1995). In this instance, the softpredecessor-

coordination-relationship mechanism will cause

the analyst tracking agent to commit to the transmission of a

completed analyst report form to the earnings analysis agent,

which can then easily extract the portion dealing with the updated

earnings estimate.

Agent Architecture

The portfolio manager and task assistant agents have an internal

agent structure called DECAF (Distributed, Environment-

Centered Agent Framework)—a general, reusable, core agent

control architecture (Oates et al. 1995). The term architecture

here refers to the internal control structure of a single agent,

as opposed to the term organization that refers to the control

and communication structure of a group of agents.

The important features of the DECAF architecture are:

_ A set of clearly defined controlmodules (planning, coordination,

scheduling, decision-making, and monitoring) that

work together to control an agent.

_ A core task structurerepresentation that is shared by all of

these control modules. This core structure can be annotated

and expanded with all manner of details that might

be “understood” only by one or a few control modules, but

there is a core, shared representation.

Briefly, the main control functions consist of a planner that

creates or extends the agent’s view of the problem(s) it is trying

to solve, called the task structure. The coordinatornotices

certain features of that structure, and may annotate it, expand

it, communicate parts to other agents, or add scheduling constraints

to it. The local scheduler takes the rough plan and

creates a low-level schedule or schedules that fix the timing

and ordering of actions. The execution monitor takes care of

actually executing the next desired action (perhaps including

pre-emption of the action in true real-time execution).

Previous work has focussed on the design of the coordinator

and the local scheduler (Decker & Lesser 1995). Details

of the implementation of these components can be found in

the cited papers. We have extended this work to include more

sophisticated execution monitoring, using such techniques as

the TCA (Task Control Architecture) approach (Simmons

1994).

Current Status

We have already built a large part of the underlying basic organizations

and achitecture as described here. Organizationally,

our agents usually form dynamic unstructured teams using

a central “matchmaker” that accepts advertisements from

new agents about thier capabilities, queries about who might

provide certain services, and unadvertisements when agents

leave the open system. Recently we have developed “brokering”

agents that act as central points of contact for certain

types of services (centralized markets). A hybrid organizational

approach allows the use of both subforms, i.e., the use

of a matchmaker in order to find an appropriate broker. More

details can be found in (Decker,Williamson, & Sycara 1996).

Architecturally, we have developed several versions of the

basic agent internals described here. Currently, all of our agent

classes (information, task, and interface agnets) are based on

this shared architecture. Although the local scheduler is considerably

simpler than the one described in (Decker & Lesser

1995), the decision-theoretic hierarchical task network planner

is more complex and capable (Williamson, Decker, &

Sycara 1996).

WARREN, our multi-agentportfolio management system,

currently consists of 6 information agents: two stock ticker

agents using different WWW sources, a news agent for Clarinet

and Dow-Jones news articles, an agent that can extract

current and historical sales and earnings-per-share data from

the SEC Edgar electronic annual report database, another for

certain textual portions of annual reports, and of course the

matchmaker. Two task agents provide

Provide

fund.

earnings

opinion

Provide

analyst recommendations

Gather

news articles

classified

"anal. rept." Fill

out anal.

rept. form

based on news

article

update

analyst

opinion

database

get

current

earnings

get

historical

earnings

gather

new earnings

reports

gather new

earnings

expectations

for

this

company

for

other

companies

in sector

for

this

company

for

other

companies

in sector

from

industry

analysts

T

min

abstract task with quality

accrual function min

subtask relationship

facilitates relationship

Earnings

Analysis

Agent

Analyst

Tracking

Agent

Figure 3: A task structure showing one coordination relationship betweentasks in the domain of the earnings report agent and the

human analyst tracking agent.

1. a simple graphical integration of stock prices and news stories

about a single stock over time on a WWW page—the

body of the news stories can be accessed by hyperlinks, and

the information is stored persistantly so that that it survives

the task agent going off-line

2. a simple fundamental analysis of a stock with respect to its

historical sales and earnings data, along with an accompanying

graph.

Finally, a portfolio interface agent can be associated with each

user of the system. The Portfolio agent (via a WWW interface)

displays the standard information about a user’s portfolio,

allows the user to (simulate) buying and selling shares, and

displays recent pricing and news information. It also provides

access to the reports produced by the two task agents, either

continuously updated or on demand.

More information and demos can be found on the WWW

at http://www.cs.cmu.edu/˜softagents

/warren/warren.html.

Conclusions

We have presented the overall framework and design decisions

made in our multi-agent system for the management of financial

portfolios through information access, filtering and integration.

Within this framework we will explore research issues

of agent coordination and negotiation and case base structuring

for user, task and situation modeling. In addition, there

are a number of learning-related research issues we want to

explore. How do we formulate the learning task in the context

of multi agent interactions where procedural and control

knowledge must be learned? Concept learning has been

the focus of most machine learning research (e.g., (Michalski

& Tecuci 1994)). Learning of control knowledge has been

explored using case-based reasoning (e.g., (Kambhampati &

Hendler 1992;Miyashita&Sycara 1995)), and reinforcement

type learning techniques (e.g., (Sutton 1988)). This research

has been conducted almost exclusively in a single agent setting.

We want to explore strategies for multiple agent learning

of control knowledge during agent interactions. Within each

formulation of the learning task (e.g. as a case-based learning,

or reinforcement learning), there are additional more specific

issues to be explored. For example, for multiple agent casebased

learning, new case indexing and retrieval algorithms

might be necessary. In addition, the number of training cases

that must be incrementally acquired through agent interactions

for reliable learning is an open issue.

Acknowledgments

This work has been supported in part by ARPA contract

F33615–93–1–1330, in part by ONR contract N00014–95–

1–1092, and in part by NSF contract IRI–9508191.

References

Arens, Y.; Chee, C. Y.; Hsu, C.-N.; and Knoblock, C. A.

1993. Retrieving and integrating data from multiple information

sources. International Journal of Intelligent and Cooperative

Information Systems 2(2):127–58.

Collet, C.; Huhns, M.; and Shen, W. 1991. Resource integration

using a large knowledge base in Carnot. Computer.

Decker, K. S., and Lesser, V. R. 1995. Designing a family

of coordination algorithms. In Proceedings of the First International

Conference on Multi-Agent Systems, 73–80. San

Francisco: AAAI Press. Longer version available as UMass

CS-TR 94–14.

Decker, K.; Williamson, M.; and Sycara, K. 1996. Intelligent

adaptive information agents. In Proceedings of the AAAI-

96 Workshop on Intelligent Adaptive Agents.

Decker, K. S. 1995. Environment Centered Analysis and Design

of Coordination Mechanisms. Ph.D. Dissertation, University

of Massachusetts.

Etzioni, O., and Weld, D. 1994. A softbot-based interface

to the internet. Communications of the ACM 37(7).

Joh, G., and Lee, C. 1992. Stock price response to accounting

information in oligopoly. Journal of Business 65(3):451–

472.

Kambhampati, S., and Hendler, J. A. 1992. A validationstructure-

based theory of plan modification and reuse. Artificial

Intelligence 55(2-3):193–258.

Kuokka, D., and Harada, L. 1995. On using KQML

for matchmaking. In Proceedings of the First International

Conference on Multi-Agent Systems, 239–245. San Francisco:

AAAI Press.

Maes, P. 1994. Agents that reduce work and information

overload. Communications of the ACM 37(7).

Markowitz, H. 1991. Portfolio selection: efficient diversification

of investments. Cambridge, MA:B. Blackwell, second

edition edition.

Michalski, R., and Tecuci, G. 1994. Machine Learning:

A multistrategy Approach, volume IV. Morgan Kaufmann

Publishers.

Miyashita, K., and Sycara, K. 1995. Cabins: A framework

of knowledge acquisition and iterative revision for schedule

improvement and reactive repair. Artificial Intelligence 76(1–

2).

Oates, T.; Prasad,M. V. N.; Lesser, V. R.; andDecker, K. S.

1995. A distributed problem solving approach to cooperative

information gathering. In AAAI Spring Symposium on

Information Gathering in Distributed Environments.

Simmons, R. 1994. Structured control for autonomous

robots. IEEE Trans. on Robotics and Automation 10(1).

Sutton, R. S. 1988. Learning to predict by the methods of

temporal differences. Machine Learning 3:9–44.

Sycara, K., and Zeng, D. 1995. Task-based multi-agent coordination

for information gathering. In Knoblock, C., and

Levy, A., eds., Working Notes of the AAAI Spring Symposium

Series on Information Gathering from Distributed,Heterogeneous

Environments. Stanford, CA: AAAI.

Sycara, K.; Zeng, D.; and Miyashita, K. 1995. Using casebased

reasoning to acquire user scheduling preferences that

change over time. In The Proceedings of the Eleventh IEEE

Conference on Artificial Intelligence Applications (CAIA ’95).

Los Angeles: IEEE.

Sycara, K. 1989. Multi-agent compromise via negotiation.

In Huhns, M., and Gasser, L., eds., Distributed Artificial Intelligence,

volume Volume 2. Pittman.

Trippi, R., and Turban, E., eds. 1990. Investment management:

decision support and expert systems. New York: Van

Nostrand Reinhold.

Williamson, M.; Decker, K.; and Sycara, K. 1996. Unified

information and control flow in hierarchical task networks.

In Proceedings of the AAAI-96 workshop on Theories of Planning,

Action, and Control.