Intelligent Agent Technology

Intelligent Agent Technology Jeffrey M. Bradshaw Bob Carpenter Rob Cranfill Mark Greaves Heather Holmback Renia Jeffers Luis Poblete Amy Sun Applied Research and Technology Shared Services Group The Boeing Company [email protected]

Why Software Agents? ◆

Original agent work instigated by researchers studying distributed intelligence

◆

New wave of agent research motivated by two practical concerns: – Overcoming the limitations of current user interface approaches – Simplifying the complexities of distributed computing

◆

Though each of these problems can be solved in other ways, the aggregate advantage of agent technology is that it can address both of them at once: – by supplementing direct manipulation with indirect management approaches – by building in high-level, loosely-coupled collaborative capabilities “out of the box”

Evolution of System Connectivity Disjoint

Ad hoc Encapsulated

Cooperating Systems with Single Agent as Global Planner

A

Cooperating System with Distributed Agents A

A

A

A

A

Agent-Enabled System Architecture Integrated  interface to  knowledge  media Agents as  intelligent  interface  managers Agent­to­agent  communication

Interapplication communication

Agent as  personal  assistant

Agents  behind the  scenes

What is a Software Agent? ◆

Agents are software entities that function continuously and autonomously in a particular environment that is often inhabited by other agents and processes

◆

Ideally a software agent should be able to: – carry out activities without requiring constant human guidance – learn from its experience – communicate and collaborate with people and other agents – move from place to place over a network as necessary

◆

Not all software agents need be “intelligent” (agents vs. minions)

◆

There is no hard dividing line between object technology and multiagent technology

Basic Agent Characteristics Agents act autonomously to accomplish objectives. • Goal-Directed • Knowledgeable • Persistent • Proactive & Reactive

Autonomous

Agents cooperate to achieve common goals. • Communication Protocols • Knowledge-Sharing • Coordination Strategies • Negotiation Protocols

Agents adapt to their environment. • Dynamic Interaction • Alternate Methods • Machine Learning

Adaptive

Cooperative

Note: Agents can be either static or mobile, depending on bandwidth requirements, data vs. program size, communication latency, and network stability (Dyer, DARPA CoABS)

Agents and Objects Basic unit

Objects instance

Agents agent

State­defining parameters

unconstrained

knowledge, desires, intentions, capabilities,…

Process of computation

operations

messages

Message types

defined in classes

defined in suites

Message sequences

implicit

defined in conversations

Social conventions

none

honesty, consistency,…

(Adapted from Shoham)

Applications of Software Agents ◆

Office automation/engineering support – – – –

◆

Information access – –

◆

“fair” allocation of limited computing resources dynamic rerouting and reassignment of tasks

Active document interfaces – –

◆

retrieval, filtering, and integration from multiple sources Internet, intranet, extranet

Resource brokering – –

◆

mail filtering meeting scheduling intelligent assistance training and performance support

intelligent integration and presentation to suit the task dynamic configuration according to resource availability and platform constraints

Intelligent collaboration – – –

between systems among people mixture of people and agent-assisted systems

Boeing IAT Program Objectives u

More powerful agent frameworks – New KAoS release – UtterKAoS: Conversations, Security, Persistence, Mobility, Middle Agents, Planning – Incorporation of COTS components (e.g., Voyager, Java platform enhancements)

u

Easier creation of sophisticated agents – ADT, comprised initially of CDT, SDT, PDT

u

Deploy in spectrum of application areas – Current areas: Information Access, DIG-IT, NASA Aviation Extranet, DARPA JumpStart – New opportunities: Spacecraft autonomy, hybrid networking QoS, security, UCAV, engineering, manufacturing

◆

◆

◆ ◆

◆

◆ ◆

◆

Some Long-Term Requirements for Industrial-Strength Agents Architecture appropriate for a wide variety of domains and operating environments Hardware-, operating-system-, programming-languageindependent Separability of message and transport layers Foundation of distributed-object/middleware (e.g.,CORBA, DCOM) and Internet technologies Fits well into component integration architectures (e.g., ActiveX, JavaBeans, Web browsers) Principled extensibility of agent-to-agent protocol Designed to work with other agent architectures, and to allow easy “agentification” of existing software Must be able to incorporate agent interoperability standards as they evolve

KAoS Implementation Context Adaptive Virtual  Document

Component integration framework

T SGML/XML  Component

Database  Component Multimedia  Component

Agents CORBA

Local and remote databases and services

Object Request  Broker  Component  Link  Web and  tools and  Servers other  Internet  Fine­grained  services services data objects

KAoS Structure and Dynamics Birth Life

Update  Structure

Formulate/Act on   Intentions

Death

Agent Structure   • Knowledge    • Facts    • Beliefs • Desires • Intentions • Capabilities

Cryogenic   State

KAoS Extension and Generic Agent Agent Extension Specific to  Particular  Agents

Optional  Planner

Shared by  All Agents

Conversation  Support

Various  Capabilities Transport­Level  Communication Security

Generic Agent

Agent-to-Agent Communication Within an Agent Domain Agent Domain

Generic Agent Instance Agent A

Agent-toAgent Protocol

Agent B

Generic Agent Instance

Domain Manager and Matchmaker ◆ ◆ ◆

◆

◆ ◆

◆ ◆ ◆

◆ ◆ ◆

The Domain Manager: Controls entry/exit of agents within a domain, governs proxy agents (i.e., security) Maintains a set of properties on behalf of the domain administrator Provides the address of the Matchmaker to agents within its domain (i.e., naming) The Matchmaker: Helps clients find information about the location of agents that have advertised their services Forwards requests to Matchmakers in other domains as appropriate Can be built on top of native distributed object system services (e.g., trader) Agents Providing Services: Advertise their services to the Matchmaker Are notified by the Matchmaker if their services have been registered Withdraw their services when they no longer wish to provide them Agents Requesting Services Ask the Matchmaker to recommend agents that match certain criteria Are given unique identifiers for the agents that match the criteria Communicate directly with these agents for services

Anatomy of a KAoS Domain External Resource

Telesthetic Extension

Proxy to Another KAoS Domain

Mediation Extension

GA

GA KAoS Agent Domain

Domain Mgr. Extension

GA Matchmaker Extension

GA

Ext. from Foreign Domain Adapter

GA

GA = Generic Agent

Proxy Extension

GA

Conversations ◆

◆

Social interaction is more appropriately modeled when conversations rather than isolated illocutionary acts are taken as the fundamental unit of discourse Two approaches to implementing agent conversations (Walker and Wooldridge): – off-line design: social laws are hard-wired in advance – emergence: conventions develop from within a group of agents

◆

KAoS currently provides only for off-line design of conversations, represented as state-transition networks – Shared knowledge about message sequencing conventions enables agents to coordinate frequently recurring interactions of a routine nature simply and predictably. – Cohen and Smith’s semantics and joint intention theory have been used to analyze KAoS conversation policies – In the future, more sophisticated agents will either be able to use less constraining “landmark-based” conversation policies or fall back to more rigid policies with identical semantics to communicate with simpler agents – In support of this, DARPA is funding us to develop a Conversation-

KAoS Conversation Policies u u

Interaction among agents best modeled at the conversational level, rather than isolated speech acts Conversation policies are agent dialogue building-blocks that provide a set of constraints that define and restrict what can take place in individual agent conversations – Policies can be expressed via many different representation formalisms, from regular expression grammars to dynamic logics

u

Conversation policies ensure reliable communication among heterogeneous agents while lessening agent’s burden of inference – Agents choose between a greatly reduced number of possible conversational moves – Conversation manager (component of “generic agent”) assures compliance with policy; handles exceptions

u

References: http://www.coginst.uwf.edu/~jbradsha/

“Conversation for Action” Policy A->B: Request

1

B->A:Promise

2

B->A: Counter

B->A: Decline

A->B: Counter

9

7

A->B: Withdraw

A->B: Decline Report B->A: Renege

A->B: Accept

5

A->B: Withdraw

B->A: Report Satisfied

3

6

A->B: Withdraw

A->B: Accept Report

4 A->B: Withdraw

8

• Communication about commitments (promise, renege) is handled explicitly, and A can notify B when the request was not fulfilled to its satisfaction (decline report) • See formal analysis of Conversation for Action Policy in Smith and Cohen 1996 AAAI paper

KAoS Applications ◆

◆

◆

◆

◆

DIG-IT: Boeing digital data integration effort to integrate agents in next-generation PMA and BOLD NASA Aviation Extranet: Agent-assisted access to information and services over a large-scale virtual private network AHCPR CDSS Project: Long-term follow-up support for bone marrow transplant patients at the Fred Hutchinson Cancer Research Center DARPA Jumpstart Project: Development of agent design toolkit (Boeing, UWF Cognition Institute, Sun Microsystems, IntelliTek) Agents for space applications: Proposal to use KAoS for a multi-agent testbed in satellite operations, and in the development of a Personal Satellite Assistant (in preparation)

JumpStart Project Overview u

Selected under the DARPA CoABS Program – Approximately 20 other participants

u u u

Partners: Boeing , Sun, UWF, IntelliTek Collaborator: Oregon Graduate Institute (CHCC) Deliverables: – Prototype software (CDT and SDT) – Periodic technical reports and demos – Interoperability demos with other CoABS participants

DARPA’s Vision of the Future of Agents u

The Future of Agent Ensembles – Agents authored by different vendors at different times – Wide variety of agent reasoning and action capabilities – Complex operational environment: • Unpredictable universe of action • Dynamic task-specific agent teams • Collaborative, negotiated problem-solving behavior

u

The Future of Agent Developers – More agents written by domain experts; fewer agents written by agent-technology experts – Decreased ability to control agent contexts of use

Simple Agents May Not Need a Complex Theory u

Simple agent systems may require only simple models of communication to achieve their ends – – – –

u

Limited tasks, collaborations, interactions with one another Predictable all simple-agent universe of action Limited and domain-specific reasoning requirements Conversations are atomic transactions

Example: – Simple personal information retrieval agents • interact mainly with non-agent information sources • little negotiation or bargaining

Sophisticated Agents Require Sophisticated Theory u

But, consider more complex applications, involving: – – – – –

u

Higher reliability, verifiability, precision of expression Arbitrary, dynamic agent collaboration with negotiation Unpredictable universe of action Complex autonomous reasoning about other agents, plans Extensive human-agent interaction

Examples: – Electronic Commerce/Electronic Trading, Air Traffic Control, Health Care, Military, etc.

u

This requires a sophisticated multiagent communication model, e.g., conversations, with an explicit semantic foundation.

Operating in Heterogeneous “What We’ve Got Here is a Failure To Communicate” Environments

◆ ◆ ◆

Mixture of different agent frameworks Mixture of simple and sophisticated agents Approach: shared conversation and security policies, generated off-line, that increase interoperability and robustness in heterogeneous agent environments

JumpStart Technical Objectives ◆

Current Focus: Communication and Security tools

– More agents written by domain experts; fewer agents written by agent-technology experts – CP scenarios require that agent policy configuration be rapid and robust – Additional classes of development tools needed in future ◆

Help developers design reliable agent conversations – Help develop ACL semantic and pragmatic theory and standards – Provide a prototype conversation design tool (CDT) • Aid agent developers in understanding ACL semantics • Help select, specialize or generate appropriate conversation policies

◆

Help developers design reliable systems with desired agent security characteristics – Develop foundations for agent security and mobility standards – Provide prototype security design tool (SDT) allowing agent developers to easily select, specialize or generate appropriate agent security policies

Conversation Policy Example: Winograd and Flores CFA A->B: Request

1

B->A:Promise

2

B->A: Counter

B->A: Decline

A->B: Counter

9

7

A->B: Withdraw

A->B: Decline Report B->A: Renege

A->B: Accept

5

A->B: Withdraw

B->A: Report Satisfied

3

6

A->B: Withdraw

4 A->B: Withdraw

8

A->B: Accept Report

u

u

u

u u

Combining Finite-State-Based and Plan-Based Conversation Policy Approaches

Intelligent agents can use less constraining plan-based policies that give them flexibility of determining many specifics of conversational moves on-the-fly Constraints governing plan-based conversation policies make them less complicated to implement than unrestricted agent dialogue models Simpler agents will continue to rely on more rigidly defined FSM-based policies where the universe of possible moves has been pre-computed “off-line” FSM and plan-based versions of same policy must comply to same semantics and pragmatics Appropriate “version” can be negotiated between agents at runtime

Extending Semantics/Pragmatics u

Participate in ongoing ACL development – KAoS, AgentTalk, FIPA, KQML-Lite, KQML-Rite – Ultimate goal of consensus on a compositional semantics with principled extensibility

u

Analyze the ACL speech acts & conversation policies – We will study/develop basic conversation properties (e.g., the ordering, timing, sequences of communication acts) – Match representations of conversation policies to diverse levels of agent capability: • Finite-state-machine models • Landmark models • Emergent conversations

– FSM and landmark models of same policy must comply to same semantics and pragmatics; choice of model negotiated at runtime between agents – We will also investigate other pragmatic conditions imposed by context (e.g., meta-conditions on agent conversations)

CDT: An Extensible Java Toolkit for Agent Conversation Design u u

The CDT is a formal design and verification system for a given theory of agency and ACL Stanford’s OpenProof will be the core framework – OpenProof is a component-based (JavaBeans) formal heterogeneous reasoning environment • Allows development of various representations (sentences, reasoning trees, FSMs, Dooley graphs, Petri nets, etc.) • Logical fragments (deductive rules, theorem-provers) • Heterogeneous transfer rules

– Extensible to different logics and theories of agency u

Generate resultant conversation policies – Off-line design simplifies agent development and reduces burden of inference for agents at runtime – Policies mediate interaction, helping increase interoperability and robustness in heterogeneous environments

Java Security and Mobility u

u u

Java is currently the most popular and arguably the most security-conscious mainstream language for agent development Its cross-platform nature makes it well-suited for heterogeneous environments However Java 1.0-1.1 failed to address many of the challenges posed by agent software – All or nothing philosophy in “sandbox” – Lack of fine-grained resource control – Security policy implementation requires writing your own security manager – Applet mechanisms are insufficient for autonomous agent mobility

New Developments in Java Security and Mobility

u

u

Mechanisms for increasing configurability, extensibility, and fine-grained access control are under development at Sun Microsystems Java 1.2 enhancements – Applets and applications on equivalent security footings – Finer-grained configurability and better resource control – Specification of much of the security policy via an external policy file, thus separating policy from mechanism

u

These new developments provide an initial foundation for support of agent-unique requirements

Security Design Tool (SDT) u

Accelerate incorporation of required agent security and mobility features into the Java platform – Foundation of new Java security model + changes to Java VM – Work with vendors, developers, standards organizations

u

Issues for Java platform enhancement and SDT development – Agent authentication and PKI management – Secure communication – Enhanced configurability and resource management • Denial of service issues: CPU, disk, memory, display • Load balancing and grid “resource dial”

– Support for secure agent mobility u

SDT Benefits – Configurable “starter set” of agent security policies – Interoperability among different agent frameworks (grid “security dial”?) – Faster creation of robust agents by non-experts

Agent “Scram” Capabilities for Anytime Mobility

Anytime Mobility u u

Telescript provided completely transparent agent mobility Current Java-based agent systems do not – Agent system code runs inside the VM; no access to execution state

u

Advantages of transparent agent mobility – Agent code need not be structured with many entry points – Allows the agent system (as well as the agents themselves) to move agents between hosts – May be transparent to the agent (may require additional redirection of agent resources) – Supports load balancing of long running agents in the grid

u

Requires modifications to the Java VM

Airplane Troubleshooting Evolution "PAPER" BASED REFERENCE

(Standalone / Linearly Organized / Org Driven Tools (Stovepipes) / Unsync Revs / Not @ Jobsite / Rev Cycle  2 Mo.)  Today's Environment (Not Process Oriented)

DDG

AMM

FIM

CLG

Troubleshooting CBT

IPC

•  Variable Fault   Download Tools •  ACARS Reporting •  FRM •  Anecdotal •  BITE

COMPUTER BASED REFERENCE (Relevant Standalone Ref Data / Hyperlinked / Org Driven (StovePipe) / Semi­Sync Revs / @ Jobsite / Rev Cycle  2 Mo.)  PMA Prototype (Bridge to Process Oriented) •  Variable Fault   Download Tools •  ACARS Reporting •  FRM •  Anecdotal •  BITE

Dispatch Deviation Guide (DDG)

Fault Isolation Manual (FIM)

Airplane Maintenance Manual (AMM)

Illustrated Parts Catalog (IPC)

Fly or Fix

Troubleshooting

Remove/Replace Test/Restore

Parts Where & When Needed

INTELLIGENT PERFORMANCE SUPPORT & REFERENCE TOOLS (Process Based / Hyperlinked / Intelligent Agents / Multimedia / Seamless Fault Det ­ Fly or Fix Res / Rev  Cycle Š 2 Mo.) Vision (Process Oriented) Process Steps Approach

Underlying New Technologies Required

In Flight Fault Detection & Downlink

Personnel Readiness

Fault Isolation

Fly / Fix Return to Service

Update Data "Documents"

Update Airline Enterprise Data System

Electronic QRH & ACARS

Component Location Multimedia Training Enterprise Data

Agent Assisted Trouble Shooting

Multimedia (JIT) Agent Assisted R&R, Test & Return to Service

Feedback to the Fault Fix System

Feedback to Airline Enterprise System

Component Integration Architecture

Intelligent Agents

Independent Links

Wireless & Wearable Computing

Media Servers 11.130.6 Evolution

Agent Roles in Technical Information u

Agent-Assisted Document Construction At the user-interface, agents work in conjunction with compound document and web browser frameworks and document management tools to select the right data, assemble the needed components, and present the information in the most appropriate way for a specific user and situation. A

u

Agent-Assisted Software Integration Behind the scenes, agents take advantage of distributed object management, database, workflow, messaging, transaction, web, and networking capabilities to discover, link, manage, and securely access the appropriate data and services.

A

A

A

A

Aviation Extranet Goals “By the turn of the century, airlines will be able to dynamically reconfigure their flight operations for improved safety and more efficient transportation for the traveling public”

◆

Develop middleware components to integrate and extend the capabilities of aviation legacy systems on a secure extranet to support: – – –

◆

Develop Extranet Global Information Services – –

◆

Real-time aircraft and airport situational awareness and scheduling and planning functions Maintenance and operations procedures enhancements Feedback data mechanisms to design/manufacturing models and simulators

Intelligent agents Metadatabases and Data Warehouses

Conduct advanced research in decision support tools for the Aviation Community

Aviation Extranet Goals (cont.)

Meta-Dbases

Data Warehouses

Data Mining

Intelligent Agents CORBA Components Legacy Data Systems

Decision Support Systems

Extranet Network Hardware

Aviation Extranet Middleware Architecture Industry Data Sources

Industry Data Sources Maintenance/ Ancillary Meta­Dbases

Industry Data Sources Real­Time Ops Meta­Dbases

Doman Service Stations

Domain Service Design/ Stations Manufacturing

Meta­Dbases

Domain Service Stations

Intelligent Web Servers

Regulations/ Documentation Meta­Dbases

Domain Service Stations

Industry Data Sources CORBA  Interfaces

Intelligent Agents

Web Browser

●

Extranet Security

● ●

Boeing  DB 

Authenticate 

Web  Server 

Certificate  Check 

(Reverse Proxy) 

&  Certificate  Check 

CORBA  Server 

Certificate  Check 

Authenticate Once Permission­Based Access Encryptable Communication

Agent 

User  Client 

A2A (over IIOP, TCP/IP, COM) 

Agent 

HTTP 

Agent 

IIOP 

Agent 

Data Access  Agent 

Airline 

Agent 

DB  DB 

CORBA  Server 

Certificate  Check 

Agent 

CORBA  Server 

Gov't 

Certificate  Check 

DB 

CORBA  Server  DB 

Agent 

DB 

Agent-Based Framework for Information Access

Matchmaker Agent Metadata/ Ontology Agent

User Agent

User Agent

Information Broker Agent

Information Broker Agent

Information Service Agent

Information Service Agent

Information Sources

Matchmaker Agent

Information Service Agent

Metadata/ Ontology Agent

* Matchmaker is connected to almost every agent