INTENTION RECOGNITION WITH EVENT CALCULUS

GRAPHS AND WEIGHT OF EVIDENCE

Fariba Sadri

Department of Computing, Imperial College London, Queens Gate, London, U.K.

Keywords: Event calculus, Ambient intelligence, Intention recognition.

Abstract: Intention recognition has significant applications in ambient intelligence, for example in assisted living and

care of the elderly, in games and in crime detection. In this paper we describe an intention recognition

system based on a formal logic of actions and fluents. The system, called WIREC, exploits plan libraries as

well as a basic theory of actions, causality and ramifications. It also exploits profiles, contextual

information, heuristics, the actor’s knowledge seeking actions, and any available integrity constraints.

Whenever the profile and context suggest there is a usual pattern of behaviour on the part of the actor the

search for intention is focused on existing plan libraries. But, when no such information is available or if the

behaviour of the actor deviates from the usual pattern, the search for intentions reverts to the basic theory of

actions, in effect dynamically constructing possible partial plans corresponding to the actions executed by

the actor.

1 INTRODUCTION

Intention recognition is the task of recognizing the

intentions of an agent by analyzing their actions

and/or analyzing the changes in the state

(environment) resulting from their actions. Research

on intention recognition has been going on for the

last 30 years or so. Early applications include story

understanding and automatic response generation,

for example in Unix help facilities. Examples of

early work can be found in Scmidt et al. (1978) and

Kautz and Allen (1986). More recently new

applications of intention recognition have attracted

much interest.

These applications include assisted living and

ambient intelligence (e.g. Pereira and Anh, 2009,

Roy et al., 2007, Geib and Goldman, 2005),

increasingly sophisticated computer games (e.g.

Cheng and Thawonmas, 2003), intrusion and

terrorism detection (e.g. Geib and Goldman, 2001,

Jarvis et al., 2004) and more militaristic applications

(e.g. Mao and Gratch, 2004 and Suzic and Svenson,

2006). These applications have brought new and

exciting challenges to the field. For example assisted

living applications require recognizing the intentions

of residents in domestic environments in order to

anticipate and assist with their needs. Applications in

computer systems intrusion or terrorism detection

require recognizing the intentions of the would-be-

attackers in order to prevent them.

Cohen, et al. (1981) classify intention

recognition as either intended or keyhole. In the

former the actor wants his intentions to be identified

and intentionally gives signals to be sensed by other

(observing) agents. In the latter the actor either does

not intend for his intentions to be identified, or does

not care; he is focused on his own activities, which

may provide only partial observability to other

agents. Our approach is applicable to both classes,

but here we describe it for the first only.

The intention recognition problem has been cast

in different formalisms and methodologies.

Prominent amongst these are logic-based, case-based

and probabilistic approaches. Regardless of the

formalism, much of the work on intention

recognition is based on using pre-specified plan

libraries that aim to predict the intentions and plans

of the actor agent. Use of the plan libraries has

obvious advantages, amongst them managing the

space of possible hypotheses about the actor’s

intentions. But it also has a number of limitations.

For example anticipating, acquiring and coding the

plan library are not easy tasks, and if intention

recognition relies entirely on plan libraries then it

cannot deal with cases where the actor’s habits are

not well-known or if the actor exhibits new, unanti-

470

Sadri F..

INTENTION RECOGNITION WITH EVENT CALCULUS GRAPHS AND WEIGHT OF EVIDENCE.

DOI: 10.5220/0003275104700475

In Proceedings of the 3rd International Conference on Agents and Artiﬁcial Intelligence (ICAART-2011), pages 470-475

ISBN: 978-989-8425-40-9

 2011 SCITEPRESS (Science and Technology Publications, Lda.)

cipated behaviour.

The contributions of this paper are as follows.

We propose a new logic-based approach to intention

recognition based on deduction and the Event

Calculus (Kowalski and Sergot, 1986) which is a

formalism for reasoning about events, causality and

ramifications. The system we propose is called

WIREC (Weighted Intention Recognition based on

Event Calculus). It exploits any available

information about the actor and the context,

including the actor’s context-based usual behaviour,

and constraints, for example his inability to perform

certain tasks in certain circumstances.

WIREC takes into account the actor’s physical

actions as well as any knowledge-seeking actions,

and reasons with what it infers about the actor’s

knowledge. It can exploit plan libraries if any plans

correspond to the known profile of the actor, and it

can revert to a basic theory of causality if no such

plans are available or if the actor’s behaviour

deviates from his known profile. WIREC

incorporates a concept of “weight-of-evidence” to

focus the search for intentions and to rank the

hypotheses about intentions.

Chen et al. (2008) also use the event calculus for

reasoning about intentions in a framework for

assisted living, but in their work they know the

intention of the actor a priori, and use the event

calculus to plan for the intention in order to guide

the actor through the required actions. Hong (2001)

shares with us concerns about the limitations of

intention recognition based entirely on plan libraries.

In his work he does not use plan libraries and uses a

form of graph search through state changes. But his

aim is to identify fully or partially achieved goals, by

way of explaining executed actions rather than to

predict future intentions and actions.

2 MOTIVATING EXAMPLES

Example 1: A simple example for ambient

intelligence at home may be based on the following

scenario. John is boiling some water. There are

multiple possible intentions, beyond the immediate

intention of having boiled water, for example to

make a hot drink (tea or coffee), to make a meal or

to use the hot water to unblock the drain. Several

factors can help us narrow the space of possible

hypotheses and to rank them. One set of factors

involves any information about John’s usual habits

(John’s profile) and constraints, and the current

context, such as time of day, and temperature (e.g.

John usually has tea during the day if it is cold, and

he does not drink coffee).

Another set of factors involves “weight of

evidence”, which can be used if John’s profile is not

known, or in conjunction with his profile, or if John

is behaving in a way unanticipated by his known

profile. Weight of evidence can be based on what we

know about what John knows based on his

“knowledge seeking” actions (e.g. John has already

looked in the cupboard and our RFID tag readers

indicate there is no tea). It can also take into account

John’s other “physical” actions, and the accumulated

effort towards one intention or another (e.g. John

gets the pasta sauce jar, strengthening the hypothesis

that he intend to make a meal, or John takes the

boiled water to the sink strengthening the hypothesis

that he intends to pour the water down the sink to

unblock the drain).

Example 2: An example with a game flavor is as

follows. Located on a grid are towns, treasures, keys

to treasures, weapons, monsters, and agents. The

agents can move through the grid stepping through

adjoining locations, can pick up weapons, enter

towns, kill monsters, and pick up treasures. They

may have some prior knowledge about the locations

of these various entities. Each agent has one or more

intentions (goals), including killing monsters,

collecting treasures or arriving at towns. The actions

that the agents can perform have preconditions, for

example to kill a monster, the agent must have a

weapon and be co-located with the monster, and to

collect a treasure the agent must have a treasure key

and be co-located with the treasure. We have no

prior knowledge of the “profiles” of the agents. We

can guess their intentions only from their actions

(and sometimes from lack of actions), for example if

their progress through the grid gets them closer to a

weapon or to a treasure key, if they seek to move to

a grid position with a monster on it, or if despite

being co-located with a weapon they do not pick it

up.

3 BACKGROUND

The approach we take in this paper is based on the

Event Calculus (EC). This formalism has been used

for planning (Mancarella et al., 2004, for example),

and has an ontology containing a set of action

operators, symbolized by A, a, a1, a2, b, c, etc, a set

of fluents (time-dependent properties), symbolized

by P, p, p1, p2, .., q, r, neg(p), etc, and a set of time

points. There are two types of fluent, primitive and

INTENTION RECOGNITION WITH EVENT CALCULUS GRAPHS AND WEIGHT OF EVIDENCE

471

ramification.

The semantics of actions are specified in terms of

their preconditions and the primitive fluents they

initiate and terminate. Initiation, termination and

preconditions are domain-dependent rules of the

form:

Initiation:

initiates(A,P,T) ← holds(P

,T)



…



holds(P

,T)

Termination:

terminates(A,P,T) ← holds(P

,T)



…



holds(P

,T)

Precondition: precondition(A,P)

The conditions holds(P

,T)



…



holds(P

,T),

above, are called qualifying conditions. The first two

rule schemas state that at a time when P

, …, P

hold, action A (if executed) will initiate, or

terminate, respectively, fluent P. The last rule

schema states that for action A to be executable

fluent P must hold. An action may have any number

of preconditions.

Primitive fluents hold as a result of actions:

holds(P,T

) ← do(A,T

)



initiates(A, P, T

)



holds(neg(P),T

) ← do(A,T

)



terminates(A, P, T

)



= T

Ramifications hold as a result of other fluents

(primitive or ramification) holding:

holds(Q, T) ← holds(P

,T)



…



holds(P

,T)

In the rules above all the variables are assumed

universally quantified in front of each rule.

As an example of EC specification consider the

following (self-explanatory) domain-dependent

rules:

Example 3:

initiates(pushOnButton(Actor, radio), on(radio), T)

←holds(hasBattery(radio),T)



holds(neg(on(radio)),T)

terminates(pushOnButton(Actor,radio), on(radio),

T) ← holds(on(radio),T)

precondition(pushOnButton(Actor,radio),

co-located(Actor, radio))

holds(co-located(X,Y), T) ← holds(loc(X,L), T)



holds(loc(Y,L),T).

4 INTENTION RECOGNITION:

OUR APPROACH

We make the following assumptions. There are two

agents, the observer (which is the WIREC system),

and the actor, who is assumed to be a rational agent,

and may have multiple (concurrent) intentions. We

observe all the actions of the actor and in the order

they take place, and the actions are successfully

executed.

As well as actions, we also observe fluents. In an

ambient intelligence assisted living scenario, for

example, the house will have a collection of sensors,

and readings from these can periodically update the

representation of state kept by the system. Such

observed fluents will typically be properties that can

change without the intervention of the actor, for

example, whether the actor is alone or has company,

and whether it is a hot day.

An intention may be an action or a fluent. In the

former case, the actor’s actions are directed towards

achieving the preconditions of the intended action,

thus making the action executable. In the latter case

the actor’s actions are directed towards achieving the

intended fluent.

4.1 Architecture of WIREC

Figures 1 and 2 illustrate the architecture of

WIREC.

Figure 1: Architecture of WIREC.

Figure 2: Architecture of the Intention Recognizer (IR) in

WIREC.

State

Observed

actions

Hypotheses

Observed

fluents

Action

Recognizer

Intention

Recognizer

(

)

Sensor

Data

Integrity

constraints

Plan

Library

(PL)

Basic

Action

Library

(BL)

Hypothesis and

weight

generator

Profile

ICAART 2011 - 3rd International Conference on Agents and Artificial Intelligence

472

The Action Recognizer can be based on some

form of activity recognition (e.g. Philipose et al.

2005), and is beyond the scope of this paper.

Observed fluents and actions update a database S

representing the current state of the environment.

The updating is done according to the semantics of

actions and fluents given by EC. S contains only

primitive fluents, the ramifications remaining

implicit. Observed actions and state S are then used

by the Intention Recognizer (IR).

IR first consults Profile to see if, in the current

context, there is any information about the actor’s

profile identifying possible intentions and plans. If

so then appropriate plans are selected from the Plan

Library PL, providing an (initial) focus for the

search. If not, or if the sequence of actions observed

thus far does not correspond to any plans that may

be selected from PL, then the search uses the Basic

Action Library, BL. Both PL and BL are based on a

graph representation of the Event Calculus.

4.2 Graph Representation of the Event

Calculus

We adopt a graph-like representation of the Event

Calculus axioms (and plans). This representation is

introduced in Table 1. Each instance of a graph

given in the last column is called a graph fragment.

This graphic representation allows our intention

recognition algorithm to be interpreted both in terms

of reasoning and in terms of graph matching or

traversal.

Plans (and thus plan libraries) can be constructed

using this graph-like representation. For example

Fig. 3(i) shows a plan for achieving intention r by

doing actions a1, a2, a3 in any order, and doing a4

after a1 and a2. Fig. 3(ii) gives a more conventional

representation of the same plan used by other

intention recognition systems. The approach in Fig.

3(i) compared to Fig. 3(ii) and to other approaches

such as the Hierarchical Task Network models (Erol

et al., 1994) has a number of advantages.

The representation in Fig. 3(i) provides

information about qualifying conditions (p1 and p2

for the initiation of q1), preconditions (q1 and q2 for

the executability of action a4) and ramifications (r

holding as a result of r1 and r2). All this information

can be useful in intention recognition. For example

if the observer knows that the actor knows that p1

does not hold, then if the actor performs action a1 he

certainly does not intend q1, nor a4, and thus is very

unlikely to intend r.

Also the observer may not see actions a1 and a2

executed, but sees a4. The plan makes it clear that

a1 and a2 are needed only to establish the

preconditions for the executability of a4. So not

having observed them does not distract from the

possibility of r being an intention. The preconditions

of a4 may have already held and the actor

opportunistically executed a4.

Table 1: EC graph representation.

EC Axiom

Name

EC Axiom schema Graph

Representation

Initiation initiates(A,P,T) ←

holds(P

,T) … 

holds(P

,T)

. P

Termination terminates(A,P,T)

←holds(P

,T) … 

holds(P

,T)

. neg(P)

Precondition precondition(A,P

)

precondition(A,P

)

precondition(A,P

)

being all the precon-

dition axioms for A

. A

Ramification holds(Q,T) ←

holds(P

,T) … 

holds(P

,T)

. Q

a1 r

p1 q1

a4 r1

a2 q2 r a1 a2 a3 a4

a3 r2

3(i) 3(ii)

Figure 3(i): An EC plan for achieving an intention r

3(ii): A conventional representation of the plan.

4.3 Generating Hypothesis

by Graph Traversal

Plan libraries (PL) in WIREC consist of “joined-up”

graph fragments such as the one in fig. 3(i), and

basic action libraries (BL) in WIREC consist of

graph fragments such as instances of those in table

1. Whether the intention recognizer uses PL or BL,

the search for hypothesis about intentions focuses on

the executed actions, propagating them through

graph matching (which can also be thought of as

forward reasoning) and propagating the “weight of

INTENTION RECOGNITION WITH EVENT CALCULUS GRAPHS AND WEIGHT OF EVIDENCE

473

evidence”. Weight of evidence, which is a number

between 0 and 1, takes into account several factors,

amongst them how many actions the actor has

executed so far towards an intention, and what the

actor knows because of his knowledge-seeking

actions.

Note that when the search uses BL, it amounts to

dynamically constructing new partial plans matching

the executed actions. We illustrate the algorithm by

an example.

Example 4: Suppose BL consists of the fragments in

table 2, where a,b,c,d,e are actions, and p, p1, p2,

p3, q, q1, .., q4, r, r1, t are fluents. Fragment 2i and

2iii represent action preconditions, 2viii represents a

ramification and the others represent fluent

initiations.

Table 2: An Example of Part of BL.

2i 2ii 2iii 2iv 2v

p a a q q b c p1 a t

2vi 2vii 2viii 2ix 2x

b q1

b q2

q2 r

d r1

e p3

Suppose we observe that action a has been

executed. Reasoning forward from a amounts to

traversing (some of) the paths starting at a. We

assign weights as we do the traversal: <q,1> and

<t, 1> (because of 2ii and 2v, q and t actually hold

now because of a), <b,1> (2iii, action b is enabled -

i.e. its precondition(s) now hold because of a),

<q1,1/2> (2vi, action b is enabled by the actor but

he has made no effort towards p1 yet, so only one

half of the conditions for achieving q1 are in place),

<q2,1/2> (2vii, similar to 2vi), <r,1/4> (2viii, the

actor has made some effort towards q2 but none

towards q3 yet), <p3, 1/4> (2x, similar to 2viii).

Notice that we ignore 2i, 2iv, 2ix; this is because

we focus on the changes that are brought about by

the actor. Now suppose the actor does action c next.

This gives a weight of 1 to p1, and increases the

weights of q1 to 1, and p3 to 1/2. The other weights

remain the same.

Our approach has a flavour of GraphPlan (Blum

and Furst, 1997), but with two significant

differences. Firstly in GraphPlan in each state all

actions whose preconditions are satisfied are

considered. In our approach we consider only those

actions whose preconditions are (fully or partially)

satisfied because of the actor’s actions. Secondly

GraphPlan completely constructs all states as it

computes paths into possible futures. We simply par-

tially “skim” paths into the future.

4.4 Controlling the Search

for Hypotheses

We make use of several features to control the

search for hypotheses:

(1) Intentions versus Consequences: An action can

have several effects, some of which may be

incidental and side-effects (e.g. increasing the water

thermostat increases the heating bill). These we call

consequences. Other effects may be the (immediate)

intentions behind the execution of the action (e.g.

having hot water) and possibly paving stones

towards further actions and longer term intentions

(e.g. having a shower and going to work). We use

consequences to update the state S, but we ignore

them in the graph traversal.

(2) Integrity Constraints: We represent and use any

available information about what the actor is not

capable of doing (e.g. he cannot climb ladders), as

well as any constraints known about the

environment (e.g. it is not possible to open the attic

door or it is not possible to enter a room without

being seen by a sensor). Integrity constraints can be

context-dependent (e.g. John never watches TV

when he has company). In example 4 if we can infer

that action e is not possible for the actor then we

will ignore fragment 2x and any other paths

originating from e.

(3) Weight of Evidence Heuristic Threshold: We

specify cut-off points, beyond which the Intention

recognizer does not look further into possible

futures. Currently we use a numerical Threshold,

such that when the weight of a fluent/action falls

below it no further reasoning (propagation) is done

using it.

(4) Knowledge-Seeking Actions: Observing the

actor’s knowledge-seeking actions (e.g. opening and

looking inside a cupboard) gives us information

about what he knows. This information is used to

increase or reduce weight of evidence of the

hypotheses, thus affecting further graph traversals,

sometimes cutting off propagation, much as with

integrity constraints.

(5) Profiles: This includes any information available

(or acquired through learning) about the actor’s

usual behaviour in given contexts, in terms of what

his intentions may be and how he may go about

achieving them. We use such information to

highlight plans in PL to focus the search on.

Typically PL’s plans are “connected” subsets of the

ICAART 2011 - 3rd International Conference on Agents and Artificial Intelligence

474

(6) fragments in BL, and thus they allow more

efficient (but less exhaustive) search for hypotheses

about intentions.

5 IMPLEMENTATION

We have a prototype implementation of WIREC,

which has been tested by test corpora generated

automatically via the planning functionalities of the

Event Calculus. We have conducted empirical

studies regarding the impact of factors (2), (3), and

(5) from the list in sub-section 4.4. The tests have

confirmed expectations regarding reduction of

search. However, larger scale tests and a realistic

application are needed and are part of future work.

6 CONCLUSIONS

In this paper we proposed an approach to intention

recognition based on the Event Calculus. The

approach has been implemented and we are currently

conducting systematic testing and empirical studies

in performance and scalability.

WIREC allows many further extensions and

enhancements, amongst them a more sophisticated

notion of weight of evidence, possibly combined

with probabilities, as well as extensions to deal with

scenarios involving partial observability or

cognitively impaired actors, or groups of actors.

Formal analysis of complexity and soundness of the

approach are also subjects of current research.

REFERENCES

Blum, A. L., Furst, M. L., 1997. Fast planning through

planning graph analysis, Artificial Intelligence, 90,

281-300.

Chen, L., Nugent, C., Mulvenna, M., Finlay, D., Hong, X.,

Poland, M., 2008. A logical framework for behabviour

reasoning and assistance in a smart home,

International Journal of Assistive Robotics and

Mechatronics, Vol. 9 No.4, 20-34.

Cheng, D. C., Thawonmas, R., 2004. Case-based plan

recognition for real-time strategy games. In

Proceedings of the 5th Game-On International

Conference (CGAIDE) 2004, Reading, UK, 36-40.

Cohen, P. R., Perrault, C. R., Allen, J. F., 1981. Beyond

question answering. In Strategies for Natural

Language Processing, W. Lehnert and M. Ringle

(Eds.), Lawrence Erlbaum Associates, Hillsdale, NJ,

245-274.

Erol, K., Hendler, J., Nau, D., 1994. Semantics for

hierarchical task-network planning. CS-TR-3239,

University of Maryland.

Geib, C. W., Goldman, R. P., 2001. Plan recognition in

intrusion detection systems. In the Proceedings of the

DARPA Information Survivability Conference and

Exposition (DISCEX), June.

Geib, C. W., Goldman, R. P., 2005. Partial observability

and probabilistic plan/goal recognition. In Proceedings

of the 2005 International Workshop on Modeling

Others from Observations (MOO-2005), July.

Hong, Jun, 2001. Goal recognition through goal graph

analysis, Journal of Artificial Intelligence Research

15, 1-30.

Jarvis, P., Lunt, T., Myers, K., 2004. Identifying terrorist

activity with AI plan recognition technology. In the

Sixteenth Innovative Applications of Artificial

Intelligence Conference (IAAI 04), AAAI Press.

Kautz, H., Allen, J. F., 1986. Generalized plan

recognition. In Proceedings of the Fifth National

Conference on Artificial Intelligence (AAAI-86), 32-

38.

Kowalski, R. A., Sergot, M. J., 1986. A logic-based

calculus of events, In New Generation Computing,

Vol. 4, No.1, February, 67-95.

Mancarella, P., Sadri, F., Terreni, G., Toni, F., 2004.

Planning partially for situated agents, 5th Workshop

on Computational Logic in Multi-Agent Systems

(CLIMA V), 29-30, September, J.Leite and P.Torroni,

eds, 132-149.

Mao, W., Gratch, J., 2004. A utility-based approach to

intention recognition. AAMAS Workshop on Agent

Tracking: Modelling Other Agents from Observations.

Pereira, L. M., Anh, H. T., 2009. Elder care via intention

recognition and evolution prospection, in: S. Abreu,

D. Seipel (eds.), Procs. 18th International Conference

on Applications of Declarative Programming and

Knowledge Management

(INAP'09), Évora,

Portugal, November.

Philipose, M., Fishkin, K. P., Perkowitz, M., Patterson, D.

J., Hahnel, D., Fox, D., Kautz, H., 2005. Inferring

ADLs from interactions with objects. IEEE Pervasive

Computing.

Roy, P., Bouchard B., Bouzouane A, Giroux S., 2007. A

hybrid plan recognition model for Alzheimer’s

patients: interleaved-erroneous dilemma. IEEE/WIC/

ACM International Conference on Intelligent Agent

Technology, 131- 137.

Scmidt, C., Sridharan, N., Goodson, J., 1978. The plan

recognition problem: an intersection of psychology

and artificial intelligence. Artificial Intelligence, Vol.

11, 45-83.

Suzić, R., Svenson, P., 2006. Capabilities-based plan

recognition. In Proceedings of the 9

International

Conference on Information Fusion, Italy, July.

INTENTION RECOGNITION WITH EVENT CALCULUS GRAPHS AND WEIGHT OF EVIDENCE

475