Capturing Expert Knowledgefor BDI-Based Behaviour Modelling

Emma NorlingCentre for Policy Modelling

Based on a portion of my PhD work @ the University of Melbourne“Modelling Human Behaviour with BDI Agents”

Outline

• Motivation• Background: BDI and models of expert

behaviour• Some approaches to knowledge acquisition• A little more depth on a couple of approaches• Tailoring techniques for BDI-based modelling• Partial worked example• Knowledge acquisition for other frameworks

Motivation

• Military simulation– Detailed models of individual behaviour– Subjects acting as experts in highly specialised

domains– Range of purposes for models, including tactics

analysis, equipment analysis, logistics planning• BDI-based models of behaviour already used

extensively

What is BDI?

• Beliefs – Desires – Intentions model of agency– Based in folk psychology, the way we think we

think

AGENT

Goals

Beliefs

Intentions

Sensors Actions

Plans

Reasoner

Goals

Expert Behaviour

• Knowledge acquisition and encoding is a difficult task

• Modelling expert behaviour has additional complications:– Much of the expert’s knowledge is “second

nature” to the expert– Person performing knowledge acquisition often

has little or no experience of the task

Sources of Inspiration

• Expert Systems• Participatory Modelling• Cognitive Task Analysis• Other subject-based techniques:– Personality profiling (Barry Silverman et al, U. Penn)– Activity traces (Georgeon et al, INRETS France)– Few published details

Applied Cognitive Task Analysis (ACTA)

• Originally developed as a means of gathering task knowledge for the development of training programs and for task redesign

• Series of semi-structured interviews1. A preliminary interview, to develop a task diagram.2. A knowledge audit, contrasting expert and novice

behaviours.3. One or more simulation interviews, resulting in mental maps

of the incidents.• Purpose is to develop a picture of the whys behind the

task, as opposed to a procedural description

Critical Decision Method (CDM)

• Aim is to capture expertise by exploring an incident that displays expertise in action

• Four “sweeps”: 1. Incident verification and selection.2. Time line verification and decision point

identification.3. Deepening, and “the story behind the story.”4. “What if...” queries, expert/novice differences,

decision errors, etc.

A Tailored Approach1. Direct observation, serving three purposes:

a) Pre-filtering the subjects to ensure they are suitable for the modelling,

b) Preparation of the interviewer with respect to the domain, and c) Providing reference points for discussion in later interviews.

2. Development of a task diagram, as in ACTA3. An iterative process of:

a) Expansion of each of the elements in the diagram through a combination of

i. probes as per the knowledge audit process of ACTA,ii. presentation of hypothetical situations, and

iii. CDM sweeps using scenarios proposed by the subjects.

b) Analysing the data and designing the models.

Example

• Work was motivated by modelling human behaviour in the military domain– Access to subjects and scenarios in this domain is

difficult• Quake 2 provided a readily-accessible domain

with an abundance of willing subjects for modelling

Developing a Task Diagram

• When you play the game, do you perceive any distinct phases? • What are your main goals in each of these phases? • What are the relative priorities of these main goals? • Say you enter a new game, where you don’t know the world map, and you

may know some, but not all, of the players. What are the first things that you do?

• Do you make an effort to get to know the style of the other players? How do you use that knowledge?

Expanding the Task Diagram• You say the first stage of the game is when you don’t know the map. When do

you consider that you do know the map? Do you explore every nook and cranny?

• What makes a good sniping spot? • If you’d just respawned and you could hear but not see a fight nearby, what

would you do? • How important are the sounds in the game to you? What sorts of things do

you listen for? • What sort of things most clearly differentiate novice players from expert

players? • Say you’d identified a particular opponent as being a better player than you.

Would you make an attempt to actively avoid him/her? • When you were talking about exploring the map, you said you explore

cautiously. What do you mean by cautiously?

Analysing the data

• Aim was to identify1. The goals of the subject, 2. The plans used by the subject, and 3. The beliefs of the subject, or more accurately,

the things about which the subject had beliefs.

Interview ExcerptSubject: ...hearing doors open is a good one, or hearing lifts activate. Like for instance if you’re up the top on that level you can hear the lift coming before it gets there, so...Interviewer: Right, so you’re ready to shoot at them? Subject: As soon as you hear the lift activate, you can start lobbing grenades into it....Subject: One of the things that it’s common to do on this level... and... a lot of other levels, is to run across the lift, activate it so it comes up... umm... people will start... lobbing grenades in. Then if you’re quick enough you can actually run around the corner and umm... beat the lift to the top. And... shoot them as they’re lobbing grenades into the lift.

Leading to…

Further Iterations

• “You said you used sounds like doors opening and lifts activating. Which other sounds do you pay attention to?”

• “With your trick of activating a lift and then trying to run around and surprise a person lobbing grenades in, do you just assume that there will be someone there, or do you use other information to decide when/if to do this?”

• “If there was more than one route that would get you to the top of the lift in time, how would you decide which one to use?”

More Generally…

• This approach was designed particularly to complement the folk-psychological underpinnings of BDI

• Also, it was designed to look at detailed models of individual behaviour

• Could it be used more generally?– Yes, but probably still suited to detailed models– There is a difficulty translating human knowledge

into code, caution must be used!