Modelling Recognition-Primed Decision Making Under Time Pressure and Multiple Situational Contexts

Copyright © IFAC Man-Machine Systems. Kyoto. Japan. 1998

MODELLING RECOGNITION-PRIMED DECISION MAKING UNDER TIME PRESSURE AND MULTIPLE SITUATIONAL CONTEXTS

Tetsuo Sawaragi' and Osamu Katai"

• Dept. of Precision Engineering, Graduate School of Engineering, Kyoto University Yoshida Honmachi, Sakyo, Kyoto 606-8501, Japan . •• Dept. of Systems Science, Graduate School of Imformatics, Kyoto University Yoshida Honmachi, Sakyo, Kyoto 606-8501, Japan.

Abstract. In this paper, instead of designing the automation system that is intended to replace the human operator completely with it and to exclude him/her out of the loop, we introduce an idea of interface agent as a sophisticated associate for a human operator. Since its decision making style must be close to the human's proficient naturalistic decision making, we formalize its resource-bounded reasoning under critical time pressure by joining a machine learning method for concept induction and a classical decision theory. Copyright © 1998 IFAC Keywords. Man-machine systems, decision support theory, machine learning, interfaces, human-centered design.

In the following of this paper, we at first introduce an idea of interface agent as an alternative to the conventional stand-alone, isolated automation systems. Then, we will characterize that such an interface agent's reasoning tasks and its decision making styles are quite different from the ones of conventional expert systems and from conventional decision support systems in that they are severely bounded to the realtime contexts. This is close to the human's naturalistic decision making style called recognition-primed decision making (Klein, 1993). We present a methodology for designing such an interface agent and formalize its resource-bounded reasoning under critical time pressure by joining a machine learning method and a classical decision theory.

1. INTRODUCTION

Recent popular concept of human-centered automation design has stressed the importance of the design philosophy of "people are in charge" or "human-in-the-loop" (Rouse, 1988). For this idealistic and very broad concept, however, Sheridan presented ten alternative meanings with a restrictive qualification saying that the real potential is somewhat questionable (Sheridan, 1997). As Sheridan depicted, we would like to stress that a humancentered design principle has to be able to answer to the following critical issues; • how to make the human-autonomy and the mechanical autonomy (i.e., automation systems) coexist letting them keep a friendly and sharable partnership • how to avoid the human from the flood of data from the plant and of comp-uter-based advice • how to allow the human operator's response variability for enhancing and maintaining his/her proficient skill levels and learning

2. INTERFACE AGENT AS A HUMAN ASSOCIATE An interface agent is a semi-intelligent computer programs that can learn by continuously "looking over the shoulder" of the user as he/she is performing actions against some complex artifacts and is expected to be capable of providing the users with adaptive aiding as

227

well as of alternating the activities instead of a human (Maes, 1993; Maes, 1994). In this sense, an agent has to coexist with a human user so that it can evolve by itself as a human user's proficient level improves. It also has to be able to stimulate a human user's creativeness coordinately by changing its role dynamically as a human's associate, rather than to replace the human user with itself.

"_"_~~m

anomalous states of plant

[41

-..~.~.~.:~~!.!~~~!~~.~......: formed from simulation data (off line)

o: :6 -~ .r-: dJ

human operator

user/operator in a time-critical situation like at an emergency. This is schematically illustrated in Fig.1. Herein , different from the conventional supervisory agent 's task of seeking for optimizing the isolated control task, such an agent has to be able to maintain its identities as an organism living within the multiple contexts looking inwards to consider the the nature of memory and perception, and looking outwards to consider the nature of social action with a human operator.

3. DYNAMIC CATEGORIZATION OF PLANT ANOMALIES

interface agent

r1\ HIC

~

human's

autonomous

'--_ _ _ _ _ CO_IIa_bora_t_or...J11 problem mental states

~Iver

,.r,~.:


Cl.i~ :~

c=J TIC

Important aspect of our interface agent is how to organize an "appearances" of the world (i.e., a complex plant), which is different from a mere collection of objective features but is specific to a particular human operator depending upon current time-criticality he/she is forced to work. To be a human-friendly associate for a human operator, an interface agent has to be able to present the current status of the plant so that he/she can help his/her situation awareness and can remind him/her of the recovering operation to be adopted. In time-critical, high-stakes situations, the time required by people to review information, and confusion arising in attempts to process large amounts of data quickly, can lead to costly delays and errors. Numerous psychological studies have provided evidence that human's decision quality may degrade with increases in the quantity and complexity of data being reviewed, and with diminishment in the time available for a response, while the rate of his/her performing tasks can be increased by filtering or suppressing i'r relevant information.

H~'!~ .

"~'k.

actions A

Fig. 1. An interface agent situated within the humansystem interactions. As for the design of the interface agent, we need to introduce some novel principles different from the conventional orientation. Currently there occurs a rapidly developing paradigmatic shift from the conventional rationalistic orientation toward an ecological view that the proper units of knowledge or sources of intelligence are primarily concrete, embodied, lived, concentrating on how to deal with the domain of immediate present where the concrete actually lives (Winograd and Flores, 1986; Lave, 1988; Suchman, 1977). A similar paradigmatic shift is also ongoing in the field of decision research (Klein et aI., 1993); a shift from a classical normative decision making paradigm toward a naturalistic decision making paradigm. The latter has concentrated increasingly on the proficient experts' situation assessment ability and their ways of looking at a situation and quickly interpreting it using their highly organized base of relevant knowledge. That is, these are abilities to recognize and appropriately classify a situation. Hereafter, we call this style of decision making as a recognition-primed decision (RPD) model after Klein (Klein, 1993).

Fig.2 shows the situation under which an interface agent is supposed to work. This illustrates a particular type of anomaly "shutdown of ANG-lines" that is generated by a simulator for a gas production plant and is used in our previous publication (Sawaragi et aI., 1996a) . This shows that the anomaly has occurred and the observed trend data begin to change from the normal states, An interface agent encountering such an anomalous situation has to be able to do the following as well as to assist the human operator to do the following; to identify the occurring anomaly type, to predict the consequences , and to determine and execute the appropriate recovery operation. Moreover, these multiple tasks must be finished until the deadline comes, or in more severe cases, they must be done as soon as possible, wherein value of the action effectivity may be degraded as the time elapses.

The distinguishing feature of the RPD model is to attempt to describe how people bring their experience to bear on a decision and how they are able to continually prepared to initiate workable, timely, and cost effective actions. Especially we are interested in modeling their capability to act proficiently under severe time pressure (i.e., under emergency); to identify the situation quickly and accurately and to act promptly with less time and effort to act.

Under such a time-critical situation, an interface agent must be able to flexibly organize the appropriate appearance of the plant status discriminating among what is now relevant and what is not for assisting an operator's situation awareness. In our system, a taxonomy of all possible plant anomalies is organized in a hierarchical fashion using a machine learning technique called concept formation (Fisher, 1986) as shown in Fig.3 (Sawaragi

In the following of this paper, we investigate into a recognition-primed decision model of an interface agent , that is embedded inside the human-artifact interactions and has to work as an intelligent associate for a human

228

et al. , 1996a) . Wherein , the root node represents a class of concepts covering all possible anomaly types , and the leaf nodes represent the individual anomalous cases (i.e., a hierarchy via is-a and subset-of relations) . In terms of this hierarchy, determination of the appearance of the plant concerns with how to determine the appropriate categorization of the plant anomalies out of the hierarchical taxonomy (Fig.3) ; to find a set of exclusive subtrees whose extensions are regarded as equivalent . The category of the hypothesis on the plausible anomaly types can be defined with a variety of abstraction levels of the hierarchical taxonomy. We call each possible categorization within the taxonomy a conceptual cover, that is a categorization of all possible anomalies into mutually exclusive and exhaustive classes.

probabilistic reasoning model (i.e. , an infiuence diagram (Howard and Matheson, 1983)) as an agent 's decision model that derives an appropriate action inferring the most plausible plant anomaly type by getting partially observable symptoms from evidential observations obtained so far. This is illustrated in Fig.4. Wherein, a

Fig. 4. Interface agent 's resource-bounded reasoning represented by an influence diagram .

observed symptoms occurrence of anomaly

deadline

\-.

t----.=--,-..."..~)

IkM~~~~~ "" " ""

4

~~

.

chance node H denotes a hypothesis node representing a state of the plant taking the value of anomaly types. Nodes Ei'S represent a set of symptoms available to an interface agent , each of which corresponds to the features observable in individual instrumentations of the plant . As knowledge of plant anomalies, an agent has acquired dependency among the classes of anomaly types and their associating evidences as well as the prior probabilities of occurrence of those hypothetical classes through the concept formation process, all of which are stored in a taxonomy hierarchy (Sawaragi et al. , 1994). Based on this apriori knowledge, getting new observations from the plant, an agent updates its prior belief on probable anomaly types based on a Bayes theorem . Then, these posterior beliefs are used for calculating the expected utilities of the available options defined in the decision node A based on the knowledge that are defined in the diagram defining the interrelationships among nodes of H , A, T and V. The option having the maximum expected utility would be determined as a recommendation for an interface agent to adopt. A diamond node is a special type of an oval node and is called a value node representing an agent's comprehensive utility. This shows that an agent 's utility is determined by the types of anomaly and the adopted action as well as by the delay. This reflects the fact that some anomaly types require an immediate recovering operations as soon as possible and if it is delayed its utility would drastically decrease (i.e., a time-dependent utility (Horvitz, 1991)} .

==

IkM~BNG

flow oontrot of ANG

<

.

-'--'-'--'-'- -

•

I

~PG flow control

- - --- ,-.-'-,- -

I

I

i'.

elapse of time

. . ':' .. ~--t...

time available to an lA

Fig. 2. An example of plant behaviors for an anomaly "shutdown of ANG-lines" . root level· ···· ··· ····· ·· · ·· ···· ··

level

1· · ·· · · ······ · · · ···· ~,..;-......

conceptual cover Ievel2· ·· ·· ··· · · · · ·-~~~~~~-

level 3 .. ..

consisting of 3 conceptual classes conceptual cover

consisting of 12 level 4 . ......•. •. ..... .. A++-I~~1l J Ul~~~ conceptual classes

levelS·············· ·· ·· ·· ·· .. levelS .......... . ....... ... . ....•.

Fig. 3. Conceptual cover of anomalous state concepts: a categorization adopted by an interface agent for its decision framing.

To illustrate an idea of time-dependent utility, here we assume only two anomalies for simplicity, Hl and H 2 , and two actions , Al and A 2. We can define an utility table as sown in Table 1. For instance, U(Al ' H l , to} de-

4. RECOGNITION-PRIMED DECISION MAKING MODEL

utility at time t - to Al

HI u(AI,Hj , to)

H2 U(AI ' H 2, to)

time variant

time invariant

In order to rationalize the agent 's selecting the appropriate categorization to be presented to the human operator and also to be used for its own problem solving (i.e., diagnosis) , we at first consider about a general

A2

U(A2, HI , to)

U( A2 ' H2 , to)

time invariant

time invariant

Table 1. Utility table at time t = to.

229

notes an utility obtained when an action Al is adopted immediately after the occurrence of the anomaly HI ' Actually the validity of some actions may vary while the time elapses. Here we assume that the utility of an action Al for an anomaly HI monotonically decreases according to the elapse of time as illustrated in Fig.5( a) (i.e. , U(AI , HI,t) < U(AI,HI,t O ) for to < t) , and the utilities ofu(A I ,H2,t),u(A2,HI ,t) and u(A 2,H2, t) do not vary, but are constant. Fig.5(b) and (c) illustrate how such a time-dependent utility effects on the recommendation of actions to be adopted. Here a horizontal axis represents a posterior belief on HI'S occurrence when evidences E are observed. The lines denote how the expected utilities for actions Al and A2 varyaccording to the beliefs of HI' p* denotes a belief with which actions Al and A2 are indifferent , and when HI is confirmed more strongly than p., an action Al is preferred to A 2 • Therefore, for the evidences E obtained at time tl just after to, the recommended action is Al' Fig.5(c) shows the similar relations after some time has passed. At this time t = t2, u( AI, HI, t2) decreases, so the indifferent belief p. shifts towards the right. Consequently, the recommended action for the same evidence E as the ones at time tl turns out to be A 2 • In our system, we categorize the utilities defined for all the combinations of hypothesis and actions into one of the three classes of time-dependent utilities as shown in Fig.5(a), each of which corresponds to the utility for the hard-deadline, the utility for the soft-deadline and the time-invariant utility, respectively.

utility

u(AI. H2.

L -_ _ _..L.___' - -_ _ _ _ _ _...J

0.0

p.

1

P(HI)

U(Al. H,. 11)

1.0

P(HI I E..~) at time u (b) utility U(Al.Hl. t2) U(AI, H,.U) u(AI . H,. 11)

1

u(AI.H,.u) L -_ _ _-L.---!_-L._ _ _ _ _...J U(Al.

0.0

p.

---+--

p. p(HI I E..~ ) at time U

H,. u)

P(HI) 1.0

(c)

Fig. 5. (a) A time-dependent utility. (b) Expected utilities for two actions at time t = t l . (c) Expected utili ties for two actions at time t = t2 (> t I > to)· types making up the conceptual cover z, and p(HJ I E,O is the probability over them, given observations E and background state of information ~. Note that for each conceptual covers z, a corresponding decision model may be constructed by retrieving the quantified conditional dependencies that are learned in the concept formation process and by joining them into a model.

5. AGENT'S MANAGING COMPLEXITY UNDER TIME-CRITICALITY As an intelligent associate, an agent aims to support an operator to focus his/her attention into the actual candidates of the anomalies rather than enforcing him/her execute a recommended action. That is, the agent presents the appropriate conceptual cover highlighting a set of possible anomalies while hiding its competitive other classes of anomalies behind (i.e., distinction between "figures" and" grounds" in psychological terms) . At the same time, this conceptual cover makes up the agent's decision models (i.e., a hypothesis node H) that calculate the beliefs on the occurrence of possible anomalies based on the currently available evidences.

(1)

and

= 2: u(A;, HJ)p(HJ I E, ~),

Al

U(AI. H2. tl) I-;'--"::::::""~~_____ I AI

j

EVC(z)

utility

utility U(A2.Hl. t,j

Given a particular conceptual cover z, we denote the recommendation obtained by solving a model constructed using this by A;, and define an expected value of categorization w.r.t. a conceptual cover z as an expected utility calculated for the recommended option A; and denote this by EVC(z) . Denoting a set of symptoms as a vector E and neglecting the changes of utilities according the elapse of time, A: and EVC(z) can be represented as follows;

2: U(Ai, HJ)p(HJ I E,O

1-_ _ _--"(lc:::im::.c•.::;in~vilf1:.:;·a::.:nl:.:..)_ (hard·deadline)

In terms of this decision model, a definition of a domain of a hypothesis node corresponds to a conceptual cover showing an appearance of the plant to an interface agent. This definition reflects the granularities of an agent's recognizing anomaly classifications, so we call this style of decision making as a recognition-primed decision making. The qualities of each possible conceptual cover must be evaluated in terms of the effect ivity brought about by the adopted actions (i.e., expected value of categorization (Poh et al., 1994». This can be formulated as follows.

A; = argm:x

tl)

(2)

j

where Hj represents a set of hypothesis classes of anomaly

230

Here, we set the following criteria concerning with the determination of the appropriate conceptual cover:

cover z and denote this by NEVC(z), which is defined using the time-dependent utility as follows 2 ;

(1) An agent should take into account of the expected delay of the operator's action execution in response to the presented information . (2) An agent should aggregate anomalies into more abstract classes if the expected utilities of all possible actions for those anomalies are negligible. (3) An agent should decompose a class of anomalies into more precise classifications if their probabilities of occurrence are high and the recommended action to be adopted varies depending on the anomalies within that class. (4) An agent should avoid the risk of hiding the correct anomaly into the" grounds" and of highlighting the wrong one in the " figures " .

N EVC(z) =

L u(A~ , Hj, CT(z))p(Hj I E,O, (4) j

where CT(Z) denotes the total cost of Cd(z) and Cc(z) . In principle, an agent should consider all possible conceptual covers and pick up the one with the maximum net expected utility of categorization. In practice, however, such an approach would be highly expensive in terms of computational cost. We take an approach to this problem by iteratively changing the level of abstraction within the conceptual cover used for model construction starting with a particular conceptual cover zoo Then , to this initial conceptual cover, the following two operations are considered to either increase or decrease the level of specificity of the concepts concerned as illustrated in Fig.6. Wherein, a specialization operator is to modify a concept in the conceptual cover by replacing it with the set of its most general subclasses located in the schema hierarchy, while a generalization operator is vise versa.

The selection of the appropriate conceptual cover is important not only to present an affordance-rich information to a human operator but also to consider about an interface agent's own problem solving activity of constructing its decision making model and solving that model. At this time, an agent has to deal with tractability of decision making inference at the expense of decision quality. That is , an agent has to determine the appropriately manageable model by finding the definition of the hypothesis node as an agent 's categorization of the environment (i.e. , a plant status). Wherein, a resource-constrained agent has to be able to enhance its decisions by expending some resource to consider the tradeoff between the expected benefit of using more detailed models to increase the expected utility of action and the resource costs entailed by computing decisions with a more detailed model. Note that this activity of finding a better conceptual cover for model modification must be done by the interface agent and this is also a time-consuming activity. Therefore, an agent has to take account of the side-effects produced due to the change of conceptual cover as well as the cost performing the model refinement operation .

Given a conceptual cover zo, we will denote a model modification operation by Si and the resulting conceptual cover by Zi = Si(ZO) ' conceptual specialization

cover

conceptual generalization

Fig. 6. Conceptual specialization and generalization. Here, we can define a expected value of model modification by operation Si on a conceptual cover Zo , EV M I(Si) as the improvement of utility due to the modification;

6. FORMULATING AN AGENT'S RESOURCE-BOUNDED REASONING UNDER MULTIPLE CONTEXTS

EV M I(s;) = EVC(Si(ZO)) - EVC(zo)

Based on the above considerations , we formulate a procedure for determining the appropriate conceptual cover by explicitly introducing the time cost incurred to the delay of the operator's executing actions and to the agent's own decision making. Let us define a delay cost C d ( z) that is incurred to the time needed for the operator to respond to the presented conceptual cover z 1 . We further introduce a computational time cost Cc(z) that is incurred to the agent to solve its decision model (i.e., calculation of posterior beliefs by probabilistic reasoning) that is constructed using the conceptual cover z . Then , we extend the definition of EVC(z) towards a net expected value of categorization w.r.t . a conceptual

(5)

With the account of the change of delay cost and computational cost due to the change of conceptual cover, a net expected value of model improvement, N EV M I, can be defined analogously to the extension of EVC towards N EV C as follows ;

NEVMI(Si) = EVMI(Si) -

~CT(Si)

- C g (s i ),(6)

where ~CT(S;) = CT(Zi)-CT(ZO) represents the change in the costs of delay and computation due to the change For the simple implementation , it is usually assumed that the time dependent utility function u is of the additive form and EVC is in units of cost, i.e. , 2

NEVC(z)

1 This cost may vary according to the proficient levels of the operator. The above modeling can be s imply extended to include s uch a variation .

231

= EVC( z ) -

(Cd( Z) + Cc(z))

(3)

of conceptual cover to be adopted , and Cg(Si) is the cost performing the model modification operation. The determination of the optimal sequence of the operations can be approximated by taking a greedy, singlestep model improvement operation with the greatest N EV MI(si) until all the operations have N EV M I( si) o. In this way, an interface agent can identify the best conceptual cover.

:s

The above formulation is to reflect the agent 's deliberation tasks under multiple contexts. In time-critical, high-stakes situations, the time required by people to review information, and confusion arising in attempts to process large amounts of data quickly, can lead to costly delays and errors. Therefore, an agent should take into account of those issues, and this is essential to establishing a human-centered design . In addition to the considerations on the collaborative partner, an agent has to be also able to reason about its own multiple activities needed to perform that basic task. In other words , the situated interface agent has to be able to control itself and to reason about the self being aware of the multiple contexts or backgrounds surrounding an agent 's current task at hand.

7. CONCLUSIONS In this paper, we presented a formulation of an interface agent 's resource-bounded reasoning for displaying information to the human operator under severe time pressure. The design principle for such an intelligent artifact that is supposed to collaborate with a human must be fundamentally different from the conventional supervisory control systems. In the latter, the supervisory strategy has been dealt with as an optimization problem in some way; how to allocate the tasks to the human and the automation systems. Wherein , objects to be optimized are isolated from the optimizing agent , and it has been assumed the infinite computational resources are available to the agent. But this is not true in the design of human-machine symbiotic systems under time pressure. The agent should be embedded within the interactions with both internal and external processes , and must seek for bounded-optimality with a capability of referring to its own activities (i.e. , a self-referencing capability) ..

REFERENCES Fisher, D . (1987). Knowledge Acquisition via Incremental Conceptual Clustering, Machine Learning, 2 , pp.139-172. Horvitz, E. and Barry, M. (1995) . Display of Information for Time-Critical Decision Making, Proc. of the Eleventh Conference on Uncertainty in Artificial Intelligence, Montreal. Gibson , J.J. (1979). The Ecological Approach to Visual Perception, Houghton MifHin Company, Boston, MA.

232

Horvits, E. (1991) . Time-Dependent Utility and Action under Uncertainty, Proc. of the Seventh Conference on Uncertainty in Artificial Intelligence , Los Angels, pp.151-158. Howard , RA. and Matheson , J.E. (1983) . Influence Diagrams, in Howard , RA . and Matheson , J .E . (Eds.) , The Principles and Applications of Decision Analysis, Strategic Decision Group , Menlo Park, CA. Klein , G.A . et al. (1993a). Decision Making in Action: Models and Methods , Ablex Pub. Corp., Norwood , NJ . Klein , G .A. (1993b) . A Recognition-Primed Decision Model of Rapid Decision Making, in Klein , G.A. et al. (eds.), Decision Making in Action: Models and Methods , pp.138-147, Ablex Pub. Corp. , Norwood , NJ. Lave, J . (1988) . Cognition in Practice: Mind, Mathematics and Culture in Everyday Life, Cambridge University Press, N.Y .. Maes, P. and Kozierok, R (1993) . Learning interface agents , Proceedings of the Eleventh National Conference on Artificial Intelligence, pp.459-465. Maes, P. (1994) . Agents that Reduce Work and Information Overload, Communications of the ACM, 37-7, pp.30-40. Poh, K.L., Fehling, M.R. and Horvitz, E .J . (1994) . Dynamic Construction and Refinement ofUtility-Basec Categorization Model, IEEE Trans . of System, Man and Cybernetics , VoI.24-11 , pp.1653-1663. Rouse, W .B. (1988) . The Human Role in Advanced Manufacturing Systems, in Compton, D . (Ed.), Design and Analysis of Integrated Manufacturing Systems, National Academy Press, Washington , D.C .. Sawaragi, T. , Iwai, S., Katai , O . and Fehling, M.R (1994) . Dynamic Decision-Model Construction by Conceptual Clustering, Proc. of the Second World Congress on Expert Systems, pp.376-384, Lisbon, Portugal. Sawaragi , T ., Takada, Y., Katai , O . and Iwai, S. (1996). Realtime Decision Support System for Plant Operators Using Concept Formation Method, Preprints of International Federation of Automatic Control (IFAC) 13th World Congress, Vol.L , pp.373-378, San Francisco. Sawaragi, T. , Tani , N. and Katai, O . (1996) . Evolutional Concept Formation for Personal Learning Apprentice by Observing Human-Machine Interactions, Proc. of CSEPC (Cognitive System Engineering for Process Control), pp.122-129. Sheridan , T .B. (1997) . Human-Centered Automation: Oxymoron or Common Sense?, Proc, of IEEE Int. Conf. on System, Man and Cybernetics, Vancouver, Canada. Suchman, L.A. (1977). Plans and Situated Actions: The Problem of Human-Machine Communication, Cambridge University Press, N.Y. . Winograd , T . and Flores , F . (1986) . Understanding Com puters and Cognition: A New Foundation for Design, Ablex Pub. Corp., Norwood , N.J..