B.H. Far's Home Page

Theme 4 (old):
Qualitative Reasoning in Supervisory Control

Abstract

Qualitative Supervisory Control (QSC), a framework for knowledge-based aiding of a domain practitioner assigned to supervise a dynamic physical system, is presented. QSC meets the main function of a knowledge-based aiding supervisory control system by developing strategies to achieve control in the face of altered configuration, along with minimizing the cognitive overload. Three problems manifesting the cognitive overload, control directness, process information accessibility and counter-intuition, are identified. QSC provides solution to them through a number of qualitative techniques. Typical applications of the QSC fn an intelligent user interface, verification of the knowledge base and system diagnosis are demonstrated.

1. INTRODUCTION

The concept of control is described as the process of influencing the behavior of a dynamic system so as to achieve a desired goal [10]. Supervisory control involves a domain practitioner intermittently obtaining information from and giving instructions to a computer, which in turn continuously controls a physical process by commanding machine actuators and reading machine sensors [23]. The main function of a knowledge-based aiding supervisory control system is the ability to develop strategies to achieve control in the face of altered configuration [14], along with minimizing the cognitive overload of the domain practitioner.

Actual implementation of control laws often incorporates a substantial amount of heuristic logic [2], [15]. The practitioner is usually responsible for tuning the control system based on proper heuristics [19]. Expert system methodologies provide a practical approach to deal with heuristic logic [1], decision making activities [22], and coupling of the mathematical process model with the symbolic computing techniques [12].

The efficiency of expert supervisory control systems depends on the knowledge representation [2], mainly rule-based. While at first the rule-based approach yielded more robust behavior in terms of greater tolerance to sensor failures [3], further experience showed three man limitations [13], [23].

  1. Lack of extensive knowledge of the actual physical underlying processes.
  2. Difficulty in expressing concepts requiring more than a aingle rule.
  3. Restricted system performance by allocation of functions between the expert system and the users.

The Qualitative Supervisory Control (QSC) framework can overcome the drawbacks, through [4]:

  1. embedding a deep qualitative knowledge of actual processes in the rule base;
  2. modeling the human operator's representation and applying it to automated decision making for activation of control algorithms [12], changing the status of a physical element [15] and supervision of the whole plant.

2. SUPERVISORY CONTROL PROBLEMS

The three main sources of cognitive overload in supervisory control are depicted in Figure 1. The cognitive overload [18] due to rapid changes in large amounts of monitored data, limits the operator's performance. This may lead to cognitive tunnelling, incorrect decisions and taking inappropriate actions. A main source of cognitive overload is out-of-phase automation that brings about extra workload for interpreting and communicating with the control system. The control directness problem [11] arises when the effects of an operator's action are propagated along the causal chains until they reach the crucial target variables on which the action cannot be direct. The process information accessibility problem [11] arises when the information needed to make intervention decisions is not directly accessible and must be inferred from the data remote from the process itself.

Table 1: Cognitive overload problems
Cognitive Overload

Decision making based on:

  • Large amount of data
  • Rapid changes of data
Information accessibility:

Data for decision making is to be inferred along causal chains

Contrl directness:

Initiating actions whose effects are propagated on causal chains to a target variable

Counter-intuition:

Limitation in anticipating plant responses due to:

  • Nonliniarities in plat model
  • Neglecting influence of overlapping processes
  • long time delays in system respomses

The human operator formulates control decisions by comparing expected behavior to the desired response [2]. The counter-intuition problem arises when the ability to anticipate the plant's response becomes limited by a number of factors, mainly nonlinearities in the plant's dynamics; the mutual influence of overlapping processes and long time delays in system response.

Solution to the control directness and process information accessibility problems can be found through introducing the deep modeling and processing techniques [4]. The consequence is a reduction of the cognitive overload. Deep knowledge representation can also remove the problems with overlapping processes. The form of deep knowledge is qualitative. Qualitative deep knowledge can deal with nonlinearities in plant dynamics.

3. THE QSC FRAMEWORK

The functional blocks of the QSC framework are shown in Fig.2 [4]. This structure shows the fundamental operator's functions in a variety of tasks, comprising:

  • monitoring displayed variables due to the needs of the task;
  • assessing the situation based on the monitored information, inherent knowledge of the system operations and a knowledge of alternative predicted schemes;
  • decision-making based on the assessed situation, goals, procedures and the expected consequences of the various actions;
  • implementing the decision through altering a parameter or set point of a variable remote from the operator.
[Figure 2]
Figure 2: The QSC framework.

4. HIERARCHICAL STRUCTURE OF QSC

Three levels of abstraction in the conceptualization (Fig.2) are proposed for blending deep and shallow knowledge and control heuristics [4]:

  1. generalized level, [.]G;
  2. qualitative confluence level, [.]Z;
  3. qualitative landmark level, [.]L.

The model for all levels is an influence graph [13], in which the nodes are symbols standing for propositions, qualitative variables, landmarks, and arcs represent either causal relations or conditional qualitative operations [4].

The [.]G level represents the shallow model. Modeling primitives are propositions and logical connectives. Propositions are related by production rules and reasoning is based on truth assignment to propositions and causal propagation of truth.

In the [.]Z level, a deep model is prepared by qualitative variables and their derivatives. The modeling primitives are three-valued qualitative variables, (i.e., +,0,- landmarks) and logical connectives.

The [.]L level embodies the deepest model and has the highest precision. Modeling primitives are qualitative variables and qualitative operations. Qualitative variables in the [.] L level have higher resolution than in the [.]Z level (i.e., more than three landmark values).

In all of the available deep knowledge representation approaches, such as [12], [16], [17], [20], [21], [24], etc., transition among the levels is governed by heuristic rules. For example in the hierarchical reasoning framework [17], [24], transition is based on some heuristics such as neglecting a variable or a rapid transient. We introduce the formal rules of transition between levels in [4].

5. MODELING AND ANALYSIS TECHNIQUES

Fig.3 depicts the modeling issues. The [.]G level resembles a typical expert control system. Modeling the system by means of confluences is equivalent to introducing a deep model to the expert system. Higher precision is obtainable by modeling the system through qualitative operations.

[Figure 3]
Figure 3 Modeling issues in QSC.

Two influence graphs are used to represent the model for each level. The Ordinal Reachability Graph (ORG) [8], whose nodes are either landmarks or propositions for the [.]G and [.]Z levels, and the Qualitative Flow Graph (QEG) [5] for the [.]L level.

Fig.4 shows the qualitative analysis techniques in QSC. Typical applications of such techniques are introduced in [4], [5], [6], [7], [8] and [9].

[Figure 4]
Figure 4 Analysis techniques in QSC.

6. APPLICATION

Three main applications of the QSC are:

a. Intelligent user interface

By providing the user with an understandable explanation of the behavior of the control system mainly, interpreting and explaining the present state and understanding the instructions given by the operator, interpreting them in terms of the control policy and applying them in actual implementation, QSC can reason about the actual behavior and synthesize a desired behavior. The Qualitative Flow Graph (QEG) [5] represents the deep knowledge which embodies the Qualitative Processes (QP) of the plant and feedback controller. A version of Qualitative Simulation (QSDEDS) [6] which can encounter both continuous qualitative variables and discrete qualitative parameters, is used to derive the potential behavior of the identified processes. Each potential behavior is called a Behavioral Fragment (BF) [5]. The Relativity Analysis (RA) [5] is developed to plan appropriate control strategy and adjust the parameters for overlapping processes.

b. Verification and Validation

The Modified Causal Ordering (MCO) [7] is an efficient digraph knowledge verification technique for detecting redundancy, circularity and inconsistency in rule-based knowledge systems. It has the capability of detecting irrelevant rules in the process of truth propagation. This reduces the computation overload [7]. Qualitative Sensitivity Analysis (QSA) [8] is a technique for detecting the robustness of the rule-based system to perturbation. QSA contributes to QSC in two ways: first, it provides a solution to the control directness problem; second, it can be used to extract shallow knowledge from the deep knowledge [4], [9].

c. Fault Detection and System Diagnosis

A typical application of QSC is the qualitative model-based procedure-oriented fault detection, which has certain advantages over other techniques such as reducing the computation overload and providing sufficient transparency for the sensory input data. Two fault detection schemes are proposed: compiled (CAMP) and deep (DAMP).

CAMP is different from the symptomatic (experience-based) fault detection techniques in the sense that the diagnostic rules are generated automatically from the behavioral rules. The generated diagnostic rules are more accurate and reliable than heuristic-based diagnostic rules and can serve as part of the fast and efficient symptomatic expert diagnosis system.

DAMP allows reasoning from the interaction of the process variables and parameters at the process level, being suitable for detecting multiple and novel faults. DAMP offers off-line recording of the behaviors of the processes and their deviation, is more efficient when comparing observed behavior with recorded behavior, and is therefore useful for the real-time diagnosis.

7. CONCLUSION

The Qualitative Supervisory Control (QSC) framework for knowledge-based alding of a domaln practitioner was proposed. Three major problems in supervisory control were identified and a partial solution was outlined through some qualitative techniques.

Typical applications of the QSC in intelligent user interface, fault detection and system diagnosis were introduced. Efficient verification and validation techniques for detection and diagnosis of the damaged information were proposed.

The advantage of QSC lies in dealing with situations that conventional control falls to cover, specially the blending of supervisory and conventional control functions.

Many of the qualitative techniques introduced are in a state of flux. Although the need for their merger is felt and their absence is conspicuous, it will probably take a longer time before they fully stabilize.

REFERENCES

  1. Astrom K.J., Anton J.J. and Arzen KE.: Expert Control, Automatica Vol.22, No.3, pp.277-285, May (1986).

  2. Basu A., Majumdar A.K. and Sinha S.: An Expert System Approach to Control System Design and Analysis, IEEE Trans. SMC, Vol.18, No.5, pp. 685-694, Sept. (1988).

  3. Bernard JA. and Washio T.: Expert System Applications Within the Nuclear Industry, pp. 91-103, American Nuclear Society Press, (1988).

  4. Far 8.11.: Modeling and Analysis of Rule Based Control Systems Based on Qualitative Reasoning and Symbol Structures, PhD. Dissertation, Chiba Univ. (1990).

  5. Far B.H.: Qualitative Control: Qualitative Design and Verification of Industrial Controllers, Proc. Int. Workshop on Indust. Appi. Mach. Intel. and Vision (MIV-89), Japan, pp. 239-243, (1989).

  6. Far B.M., Nakamichi M. and Sampel M.: Qualitative Modeling and Simulation: Reasoning about Behavior of Flexible Manufacturing Systems, Proc. Int. Workshop on Computer Appl. to Production and Eng. (CAPE-89), Japan, pp. 527-537 (1989).

  7. Far B.H., Nakamichi M. and Sampei M.: Verification of Knowledge in Rule Based Systems by Modified Causal Ordering, submitted to the mt. Jour. of Man Machine Studies (IJMMS).

  8. Far B.H., Nakamichi M. and Sampei M.: Qualitative Sensitivity Analysis, Jour. Jap. Soc. for Al (JSAI), Vol. 6, No.1 (1991).

  9. Far 8.11., Nakamichi M. and Sampei M.: Automatic Diagnostic Rule Generation for Rule Based Controllers by Qualitative Sensitivity Analysis, Jour. Jap. Soc. for Electrical Eng., Oct. (1990).

  10. Fleming W.H.: Report to The Panel of The Future Directions in Control Theory: A Mathematical Prospective, Society for Industrial and Appli. Math. (1988).

  11. Moc J.M.: Strategies in Controlling a Continous Process with Long Response Latencies: Needs for Computer Support to Diagnosis, lint. Jour. of Man Machine Studies, Vol.30, pp. 47-67 (1989).

  12. Kowalik S.J. (Ed.): Coupling Symbolic and Numerical Computing in Expert Systems, Elsevier Science Publishers (1986).

  13. Kuipers B.: Qualitative Reasoning: Modeling and Simulation With In complete Knowledge, Automatica, Vol. 25, No.4, pp. 571-585 (1989).

  14. Levis A.H.: Challenges to Control: A collective view, IEEE Trans. Automatic Control, Vol. AC-32, No. 4 (1987).

  15. Mahmoud M.S., Eid S.Z. and Abou- Elsoud A.: A Real-Time Expert Control System for Dynamic Processes, IEEE Trans. SMC, Vol. SMC-19, No.5 (1989).

  16. Martinez J.M. and Nomine J.P.: Artificial Intelligence for Mission Planning for Mobile Robots, Proc. 1st Int. Conf. on Supercoraputing in Neuclear Appl. (SNA 90), Japan, March (1990).

  17. Motoda H.: Towards Deep Understanding in Problem Solving, Pros. 1st lint. Conf. on Supercomputing in Nuclear Appl., (SNA '90), Japan, March (1990).

  18. Paterson A., Sachs P. and Turner M.: ESCORT: The Application of Causal Knowledge to Real- Time Process Control, in Expert Systems' 85, British Inf. Soc. Ltd., pp. 79-88 (1985).

  19. Rasmussen J. and Goodstein L.P.: Decision Support in Supervisory Control of High-risk Industrial Systems, Automatica, Vol. 23, No. 5, pp. 663-671 (1987).

  20. Rasmussen J.: Skills, Rules and Knowledge: Signals, Signs and Symbols, and other distinctions in Human Performance Models, IEEE Trans. on SMC, Vol. SMC-13, No. 3, pp. 257-266 (1983).

  21. Rasmussen J.: The Role of Hierarchicnl Knowledge Representation In Decision Making and System Management, IEEE Trans. on SMC, Vol. SMC-15, No. 2, pp. 234-243 (1985).

  22. Sage AP. and Rouse W.B.: Aiding the Human Decision Maker through the Knowledge Based Sciences, IEEE, Trans. SMC, Vol.16, No.4, pp.511-521 (1986).

  23. Sheridan TB.: Human and Computer Role in Supervisory Control and Telerobotics: Musing about Function, Language and Hierarchy, in Task, Errors and Mental Models, Chapter 9, Taylor and Francis (1988).

  24. Yoshida K. and Motoda H.: An Approach to Hierarchical Qualitative Reasoning: Constructing Shallow Knowledge from Deep knowledge (in Japanese), Jour. Jap. Soc. for AI, Vol.4, No.4, pp.447-455 (1990).