A Cognitive Framework for Delegation to an Assistive User Agent - - PowerPoint PPT Presentation

A Cognitive Framework for Delegation to an Assistive User Agent

Karen Myers and Neil Yorke-Smith

Artificial Intelligence Center, SRI International

Overview

CALO: a learning cognitive assistant User delegation of tasks to CALO Delegative BDI agent framework Goal adoption and commitments Summary and research issues

CALO: Cognitive Assistant that Learns and Organizes

Track execution of project tasks Help m anage tim e and com m itm ents Perform tasks in collaboration w ith the user

• CALO supports a high-level knowledge worker
• Understands the “office world”, your projects and schedule
• Performs delegated tasks on your behalf
• Works with you to complete tasks
• Stays with you (and learns) over long periods of time
• Learns to anticipate and fulfill your needs
• Learns your preferred way of working

CALO Year 2

Overview

CALO: a learning cognitive assistant User delegation of tasks to CALO Delegative BDI agent framework Goal adoption and commitments Summary and research issues

Delegation May Lead to Conflicts

Focus on delegation of tasks from user to CALO

Not on tasks to be performed in collaboration One aspect of CALO’s role as intelligent assistant

CALO cannot act if conflicts over actions

Conflicts in tasks

“purchase this computer on my behalf” “register me for the Fall Symposium”

Conflicts in guidance

“always ask for permissions by email” “never use email for sensitive purchases”

Conflicts in User’s Desires

“I wish to be thin” “I wish to eat chocolate” But Richard Waldinger’s

scotch mocha brownies are full of calories

⇒ conflict between incompatible desires

User’s desires conflict with each other Humans seem to have no problem with such conflicts

CALO must recognize and respond appropriately

Other Types of Conflicts

Current and new commitments

Currently CALO is undertaking tasks to:

Purchase an item of computer equipment Register user for a conference

Now user tasks CALO to register for a second conference Set of new goals is logically consistent and coherent But infeasible because insufficient discretionary funds

Commitments and advice

User tasks CALO to schedule visitor’s seminar in best

conference room

Existing advice: “Never change a booking in the

auditorium without consulting me”

New goal and existing advice are inconsistent

The BDI Framework

CALO’s ability to act is based on BDI framework

Beliefs = informational attitudes about the world Desires = motivational attitudes on what to do Intentions = deliberative commitments to act

Realized in the SPARK agent system

Hierarchical, procedural reasoning framework BDI components in SPARK represented as:

Facts (beliefs) Intentions (goals/intentions) Desires are not represented

Procedures are plans to achieve intentions

Desires vs. Goals

Both are motivational attitudes Desires may be neither coherent (with beliefs)

nor consistent (with each other)

Goals must be both

Desires are ‘wishes’; goals are ‘wants’

“I wish to be thin and I wish to eat chocolate” “I want to have another of Richard’s brownies”

Desires lead to goals

CALO’s primary desire: satisfy its user

Secondary desires→goals to do what user asks

‘BDI’ Agents are Really ‘BGI’

Decision theory emphasizes B and D AI agent theory emphasizes B and I In most BDI literature, ‘Desires’ and ‘Goals’ are

confounded

In practice, focus is on:

goal and then intention selection

• ption generation, and plan execution and scheduling

Focus has been much less on:

deliberating over desires goal generation advisability

vital for CALO

The Problem with BGI

When Desires and Goals are unified into a single

motivational attitude:

Can’t support conflicting D/G (and D/B) Hard to express goal generation Hard to diagnose and resolve conflicts

Between D/G and I, and between G, I, and plans

Hard to handle conflicts in advice

How can CALO make sense of the user’s taskings

in order to act upon them?

How can CALO recognize and respond to

(potential) conflicts?

Overview

CALO: a learning cognitive assistant User delegation of tasks to CALO Delegative BDI agent framework Goal adoption and commitments Summary and research issues

Cognitive Models for Delegation

user agent

Buser

(do assigned tasks)

Bagent GA

Belief

Duser Dagent

Desire

GC

agent

Guser

+ + +

alignment delegation refinement decision making

Goal

goal adoption

Candidate Goals Adopted Goals satisfy all tasks

Delegative BDI Agent Architecture

user

failure conflicts revision

advice

AE AG agent

GC GA I

execute

B

sub-goaling

B D G

Candidate Goals Adopted Goals Intentions Goal Advice Execution Advice

Overview

CALO: a learning cognitive assistant User delegation of tasks to CALO Delegative BDI agent framework Goal adoption and commitments Summary and research issues

Requirements on Goal Adoption

Self-consistency: GA must be mutually consistent Coherence: GA must be mutually consistent

relative to the current beliefs B

Feasibility: GA must be mutually satisfiable

relative to current intentions I and available plans

Includes resource feasibility

Reasonableness: GA should be mutually

‘reasonable’ with respect to current B and I

Common sense check: did you really mean to purchase a

second laptop computer today?

Responding to Conflicting Desires

Goal adoption process should admit:

Adopting, suspending, or rejecting candidate goals Modifying adopted goals and/or intentions Modifying beliefs (by acting to change world state)

Example: User desires to attend a conference in

Europe but lacks sufficient discretionary funds

shorten a previously scheduled trip cancel the planned purchase of a new laptop

• r apply for a travel grant from the department

Combined Commitment Deliberation

Goal adoption

Adopted Goals ≠ Candidate Goals (≠ Desires)

Intention reconsideration

Extended agent life-cycle Non-adopted Candidate Goals Execution problems with Adopted Goals

Propose combined commitment deliberation

mechanism

Based on agent’s deliberation over its mental states Bounded rationality: as far as the agent believes and

can compute

BDI Control Cycle identify changes to mental state decide on response perform actions

world state changes commitment deliberation

Mental State Transitions

Current mental state S = (B,GC,GA,I)

Omit D since suppose single “satisfy user” desire

Outcome of deliberation is new state S' Possible new transitions:

Expansion

adopt additional goal

No modification to existing goals or intentions

Revocation

drop adopted goal + intention

To enable a different goal in the future

Proactive

create new candidate goal and adopt it

To enable a current candidate goal in the future

Plus standard BGI transitions

E.g. drop an intention due to plan failure

• bserve

decide

commitment deliberation

act

Goal and Intention Attributes

Goals:

• User-specified value/utility
• Can be time-varying
• User-specified priority
• User-specified deadline
• Estimate cost to achieve
• Level of commitment so far
• For adopted goals

Intentions:

• Implied value/utility
• Cost of change
• Deliberative effort
• Loss of utility
• Delay
• Level of commitment
• Level of effort so far
• E.g. estimated %

complete

• Estimated cost to complete
• Estimated prob. success

Making the Best Decision

S→S' transition as multi-criteria optimization

Maximize (minimize) some combination of criteria over S Can be simple or complex

Bounded rationality Simple default strategy, customizable by user

Advice acts as constraints ⇒ constrained (soft)

multi-criteria optimization problem

“Don’t drop any intention > 70% complete”

Assistive agent can consult user if no clear best S'

“Should I give up on purchasing a laptop, in order to

satisfy your decision to travel to both conferences?”

Learn and refine model of user’s preferences

Example

Candidate goals:

c1: “Purchase a laptop” c2: “Attend AAAI”

Adopted goals and intentions:

g1 with intention i1: “Purchase a high-end laptop using

general funds”

g2 with intention i2: “Attend AAAI and its workshops,

staying in conference hotel”

New candidate goal from user:

c3: “Attend AAMAS” (high priority)

Mental state S = (B, {c1,c2,c3}, {g1,g2}, {i1,i2})

Example (cont.)

• CALO finds cannot adopt c3
• {g1,g2,g3} resource contention – insufficient general funds
• Options include:

1.

Do not adopt c3 (don’t attend AAMAS)

2.

Drop c1 or c2 (laptop purchase or AAAI attendance)

3.

Modify g2 to attend only the main AAAI conference

• But changing i2 incurs a financial penalty

4.

Adopt a new candidate goal c4 to apply for a departmental travel grant

• Advice prohibits option 2

Example (cont.)

CALO builds optimization problem and solves it

Problem constructed and solution method employed both

depend on agent’s nature

E.g. ignore % of intention completed No more than 10ms to solve

Finds best is tie between options 3 and 4

Agent’s strategy (based on user guidance) is to consult

user over which to do

User instructs CALO to do both options

New mental state

S' = (B', {c1,c2,c3,c4}, {g1,g'2,g3,g4}, {i1,i'2})

Overview

CALO: a learning cognitive assistant User delegation of tasks to CALO Delegative BDI agent framework Goal adoption and commitments Summary and research issues

Summary

CALO acts as user’s intelligent assistant

Classical BDI framework inadequate Implemented BDI systems lack formal grounding

Proposed delegative BDI agent framework

Separate Desires and Goals Separate Candidate and Adopted Goals Incorporate user guidance and preferences Combined commitment deliberation for goal adoption and

intention reconsideration

Enables reasoning necessary for an agent such as CALO

Implemented by extending SPARK agent

framework

Related Work

BOID framework [Broersen et al]

Different types of agents based on B/D/G/I conflict

resolution strategies

BDGICTL logic [Dastani et al]

Merging desires into goals

Intention reconsideration [Schut et al] Collaborative problem solving [Leveque and

Cohen; Allen and Ferguson]

Social norms and obligations [Dignum et al]

Future Work

Extend goal reasoning to consider resource

feasibility (in progress)

Proactive goal anticipation and adoption Collaborative human-CALO problem solving

Beyond (merely) completing user-delegated tasks

Multi-CALO coordination and teamwork Learning as part of CALO’s extended life-cycle More information: http://calo.sri.com/