You are on page 1of 5

Modeling Real-Time Decision-Making Systems With Rough Fuzzy Petri Nets

J.F. Peters1, A. Skowron2, Z. Suraj3, S. Ramanna4, A. Paryzek1


1 Department of Electrical and Computer Engineering, University of Manitoba, Winnipeg, Manitoba R3T 2N2 Canada, Phone: 204 474 9099, Fax: 204 262 4639, email: jfpeters@ee.umanitoba.ca 2 Institute of Mathematics, Warsaw University, Banacha 2, 02-097 Warsaw, Poland 3 Institute of Mathematics, Pedagogical University, Rejtana 16A, 35-310 Rzeszw, Poland 4 Department of Business Computing, University of Winnipeg, Winnipeg, R3B 2E9, Canada and Department of Electrical and Computer Engineering, University of Manitoba, Winnipeg, R3T 3E2, Canada, ramanna@ee.umanitoba.ca

ABSTRACT: This paper considers the construction of Petri nets to simulate the computation performed by real-time decision systems. A description of Petri nets which model decision-making in real-time decision systems is briefly described in Section 1. A rough fuzzy Petri net extension of rough Petri nets are presented in Section 2. These nets are used to create highly parallel programs to simulate approximate reasoning system computations. Constructed nets make it possible to evaluate the design of decision system tables, and to trace computations in rules derived from decision tables. 1 INTRODUCTION Considerable work has already been carried out in modeling decision system rules with Petri nets [Skowron, 1995; Skowron and Suraj, 1993, 1996; Peters, 1997; Peters et al., 1998]. The aim of the earlier as well as the current research has been to simplify the analysis of large information systems, and the transformation of such systems as well as derived rules into corresponding concurrent models. A complete framework for approximate reasoning is provided by rough set theory [Pawlak, 1982; Pawlak, 1991; Pawlak, 1995; Skowron, 1995; Skowron, Polkowski, 1996; Skowron, Rauszer, 1992; Skowron, Suraj, 1993a, 1993b, 1994, 1995, 1996] with extensions to real-time decision-making in [Skowron and Suraj, 1996; Peters, 1997; Peters et al., 1998; Peters, Ziaei, Ramanna, 1998]. Rough set theory provides a means of deriving rules from decision system tables [Pawlak, 1991]. Dependencies among sensor computations relative to decisions about combinations of sensor outputs are expressed as rules. Derived rules provide a basis for evaluating as well as controlling processes. Let S = (U, A 4 {d}) be a decision system with a set of objects U, set of sensors A, and decision d. A high-level flowchart of the methodology underlying both the construction of and computations performed by an approximate realtime reasoning system is given in Fig. 1. The construction of a decision system table and derivation of rules begins in state 0 and is completed with the derivation of rules in set D(S) for a system S before entering state 1. Actual approximate real-time reasoning begins in state 1, where reading inputs from experiments and selection of a timer setting t are done in parallel. Then enabled sensors evaluate inputs concurrently with the operation of an approximate time window (atw) introduced in [Peters, 1997]. The notion of a time window was introduced in [Petri, 1996]. Since there is some uncertainty in our knowledge of the exact period t of a synchronizing time window, it is appropriate to introduce approximate partitions of durations associated with a time window and obtain an approximate time window. In other words, measurements of durations between transition firings in a decision system are granulated. Next, an appropriate rule r is selected based on sensor readings. Let r = (P , (dr = v)) e D(S), where P is a premise of rule r. The notation P 0 is a shorthand which denotes the degree-of-membership of sensor readings (referenced in P) in corresponding granules. Let a e A, x e U, and let gi(ai(x)) e [0, 1] for membership function g i for sensor a i. Then P0 a means that K a e A . (gi(ai(x)) ai). In the case where sensor readings satisfy the conditions of the premise for a rule, performance of some action act(d r) governed by the value of dr is carried out. Otherwise, the system waits for new sensor readings. These actions are performed in parallel with reading the atw and adjusting a timer if necessary. 2 ROUGH FUZZY PETRI NETS In what follows, it is assumed that the reader is familiar with classical Petri nets [Petri, 1962] as well as the basic structure of coloured Petri nets found in [Jensen, 1992] and fuzzy Petri nets in [Pedrycz and Gomide, 1994]. Rough Petri nets are derived from coloured and hierarchical Petri nets as well as from rough set theory. The complete process of deriving rules for a decision system can be modeled with a rough Petri net at a sufficiently high level to facilitate an understanding of a particular rule-derivation process. In addition, the strengths of connections (weightings of attribute computations) in a rough Petri net can be calibrated. Rough fuzzy Petri nets are derived from rough Petri nets and fuzzy Petri nets. This form of Petri net makes it possible to add learning capability to a decision-making rule-derivation

process, which combines the use of fuzzy sets and rough sets in rule formulation. An overview of the types of Petri nets and related set theories leading to rough fuzzy Petri nets is given in Fig. 2. 2.1 Rough Petri Nets The motivation for introducing rough Petri nets is to capture the understandings and operations from rough set theory, which have been used to construct both general-purpose as well as highly-specialized decision systems. The complete process of deriving rules for a decision system can be modeled with a rough Petri net at a sufficiently high level to facilitate an understanding of a particular rule-derivation process. In addition, strengths of connections (weightings of sensor computations) in a rough Petri net can be calibrated.

begin

construct decision table derive decision rules in D(S) outputs rule r1 1 await inputs sensor an fire sensors set deadlinet atw clock rule rk

select rule r = (P , (dr = v)) reading atw P o a no 1 no adjust t yes yes ontime

sample measurements

granulate sample measurements

model sensors outputs sensor a1

act(dr)

Fig. 1: Basic Real-Time Decision System classical Petri nets

coloured Petri nets

hierarchical Petri nets

rough sets rough Petri nets

fuzzy sets

fuzzy Petri nets roughly fuzzy Petri nets

Fig. 2 Lineage of Rough Fuzzy Petri Nets A rough Petri net (rPn) is a structure ( S , P, T, A, N, C, E, I, W, B , p ) where S , P, T, A, N, C, E, I are as in a CPN. Strengths-of-connections (chosen from a finite set of weights W) are assigned to arcs with p : A > W. A strength of connection w i e W specifies the relative importance of input to a sensor (a form of attribute in rough sets), and guarantees a certain magnitude of input to a transition. Weights are restricted to values in the interval [0, 1]. Let X,

S, r be a set of inputs, information system S, and reduct r belonging to set of all reducts RED(S) of S, respectively. Let S = (U, A) be an information system and let R be an equivalence relation which forms the quotient set X/R, where R @ A and X @ U. Also, recall that the notation X/R (known as the quotient set) denotes the family of equivalence classes of R. The set POSR(X) = RX is the set of all elements of U which can classified as elements of X. Similarly, the set NEGR(X) = U - RX is the set of those elements of U which can be classified as elements of U - X [Pawlak, 1991]. Further, a partial list of the elements of the set B is given in (1).

X / R , POS ( X ) , NEG ( X ) , inf ( X ) , R R A dec ( X ) , M ( S ) , RED( S ) , fM ( RED( S )) , OPT ( S ) , OPT ( R, S ) A


B

(1)

contains distinguished procedures used to construct the quotient set X/R, set POSR(X) and set NEGR(X), as well as procedures to construct an information system table, decision system table, discernability matrix, set all rules for a decision system, and set of all rules relative to a reduct R, respectively. The prescription of the elements of B is nonexhaustive. The arc expression function E has been specialized relative to a finite set of rough set operations B such that E: A > B. The operations in B are used to describe processes which are the key components of an approximate reasoning system, namely, decision tables, discernability matrices and functions, reducts, and rules. Assume that each input x has a strength-of-connection w. Further, let S = ({x}, {a} 4 {d}) be a decision system with a single input, attribute a, and decision d. An example of a rough Petri net with a single transition is given in Fig. 3.

w x

decA(x) (w x, a(w x), d)

Fig. 3 Single-Transition Rough Petri Net The output of the net in Fig. 3 is a tuple (w x, a(w x), d) representing a decision system table with a single row. The computation w x aggregates w and x with the anonymous operation : [0,1]2 > [0,1]. The operation has a number of interpretations. For instance, we can define (w, x) = x + w - (w)(x) (probabilistic sum) or max(w, x) in cases where we want to check for dominance of a strength-of-connection over x. In the absence of a specific strength-ofconnection, the default value of w is 0 and w x equals x. An example of a rough Petri net is given in Fig. 4. This Petri net models the rule (va(w x) e [*, k]) , (d = vd) where vd is the value of decision d. await aS ?x x a w [a(w x) e [*, k] ] va(w x) td Fig. 4 Rough Petri Net The output of the net in Fig. 4 is vd represents a decision based on the discovery that a(w x) belongs to the interval [*, k], i.e., a(w x) k. In addition, we introduce an initialization configuration inf0, which is a set of receptive processes waiting for input from the environment. The notion of a receptive process was introduced by [Dill, 1989]. A receptive process models the interaction between a system and its environment. A receptive process must always be ready to receive input. Let ?u denote waiting for input u, where ? (pronounced query) specifies a special form of sensor. The sensor ? maps a value read from a hidden communication channel connected to the environment to an object x e U. In Fig. 4, for example, ?x specifies waiting for input x. 2.2 Rough Fuzzy Petri Nets Roughly fuzzy Petri nets were introduced as an extension of fuzzy Petri nets in [Peters, 1997]. In this section, rough fuzzy Petri nets are introduced as a straightforward extension of rough Petri nets. The introduction of rough fuzzy Petri Nets (rfPNs) is motivated by the need to develop mathematical models of decision-making systems relative to aggregations of granulated inputs such that the models are capable of learning, and are designed to react to dynamic d vd

changes in the reduct set for different samples of decision tables. A rough fuzzy Petri net (rfPn) is a structure given in (2). (S, P, T, A, N, C, E, I, M, W, B 4 Z, f , p ) (2) where S , P, T, A, N, C, E, I, W, B , p are as in a rough Petri net. The set B is augmented with operations in Z from fuzzy Petri nets to handle aggregations of granulated inputs. The modulators (also called reference points) of operations in Z are in the set M. The operation f also comes from fuzzy Petri nets such that modulators chosen from the finite set M are assigned to arcs by f: A > M. In effect, a rough fuzzy Petri net is an extension of the rough Petri net model, which provides a concise means of simulating the computation performed by an information system S. A rough fuzzy Petri net can be calibrated during supervised learning as in [Pedrycz and Peters, 1997; Pedrycz and Peters, 1998]. That is, the values of its modulators and strengths-of-connections can be learned relative to appropriate target values. For each input x in U, the operation S constructs a row in a decision system table. The results obtained for coloured Petri net constructions and simulations of information systems carry over to rough fuzzy Petri nets. The synchronization of transition firing in a decision system can be accomplished with the addition of a subnet to simulate a timer [Skowron and Suraj, 1996]. This insight is consonant with the notion of an approximate time window (measure of durations between firings of transitions in a Petri net) introduced in [Peters, 1997]. Then each decision transition is connected to a subnet (itself, a rough fuzzy net which simulates a timer decision system) which simulates an approximate time window. The advantage in doing this is that the period as well as the approximate sizes of the partitions of the synchronizing timer can be calibrated to achieve optimum performance during simulation. Assuming that the structure of the rough fuzzy Petri net model for a real-time decision system accurately reflects the operating conditions of the system, it is then possible to design intelligent system timers (i.e., approximate time windows controlled by decision system rules). The method for constructing a real-time decision system with approximate time windows is given in [Peters, 1997]. In what follows, the notation (Ns, Ms) denotes a rough fuzzy Petri net constructed for a decision table S, which is represented as the union of two disjoint sets I (net input places associated with objects of U) and V (net output places associated with a value computed by a sensor in A). A marked rfPN (Ns, Ms) is a net where each of its receptors has received a value, which initiates the construction of a row in a decision table. Again, let S = ({x}, {a} 4 {d}) be a decision system with a single input, attribute a, and decision d. In addition, we define a predicate P = ((va(x)) = k)). Then define a guard on the decision transition td relative to an identity function deg(P(a)) = P(a) such that deg(P(a)) a. Notice that deg(P(a)) is one possible interpretation of the notation P0 which prescribes an a-level in real-time decision-making in Fig. 1. A sample Petri net to simulate computations in S is given in Fig. 5. [deg(P(a)) a] await aS ?x ?start_atw start x a va(x) td d vd

datw vatw reset adjusted atw paritions

Fig. 5 Rough fuzzy Petri net to simulate S Theorem Let (m(0), ..., m(k)) be an arbitrary computation of a marked rough fuzzy Petri net net (Ns, Ms), which simulates a real-time decision system S = (U, A 4 {d}, {atw}). For each object ui in U, introduce a start place asi with matching index i. Then 1. The decision rule r in decision system D(S) is true in S provided r = {a = j: mv (deg( P( a)) ) = 1 & a A} (( d = vd ) & (d atw = v atw )), where td is the enabled transition in m
(k) (k)

2. For any j < k, it is the case that r = {a = p: mv (deg( P( a)) ) = 1 & a A} (( d = vd ) & (d atw = v atw )) is not true in S.
( j)

3 CONCLUDING REMARKS
The contribution of this paper is the presentation of an approach to modeling real-time decision systems with rough fuzzy Petri nets. The advantage to doing this derives from the fact that computations performed by such a system can be simulated and analyzed. The approach taken in this paper draws upon the extensive research on rough sets and approximate reasoning carried out by researchers in recent years.

ACKNOWLEDGEMENTS
We gratefully acknowledge the funding for this research provided by the Natural Sciences and Engineering Research Council of Canada (NSERC), Canadian Space Agency, the University of Manitoba research grants committee, and the University of Winnipeg research grants committee.

REFERENCES
Dill, D.L. 1989. Trace Theory for Automatic Hierarchical Verification of Speed-Independent Circuits. Cambridge, MA, MIT Press. Jensen, K. 1992. Coloured Petri Nets--Basic Concepts, Analysis Methods and Practical Use 1. Berlin, SpringerVerlag. Pawlak, Z 1982. Rough sets, Int. J. of Computer and Information Sciences Vol. 11, pp. 341-356. Pawlak, Z 1991. Rough Sets: Theoretical Aspects of Reasoning About Data. Boston, MA, Kluwer Academic Publishers. Pawlak, Z 1995. Rough sets: present state and future prospects. ICS Research Report 32/95, Institute of Computer Science, Warsaw Institute of Technology. Pedrycz, W.; Peters, J.F. 1997. Information Granularity Uncertainty Principle: Contingency Tables and Petri Net Representations. Proc. Proc. North American Fuzzy Information Processing Society NAFIPS97, Syracuse, NY, pp. 222-226. Pedrycz, W.; Peters, J.F. 1998. Learning in fuzzy Petri nets. In: Fuzzy Petri Nets, Cardoso, J., Sandri, S. (Eds.). Berlin, Physica Verlag [in press]. Pedrycz, W.; Gomide, F. 1994. A generalized fuzzy Petri net model. IEEE Trans. on Fuzzy Systems Vol. 2/4, pp. 295-301. Petri, C.A. 1962. Kommunikation mit Automaten. Schriften des IIM Nr. 3, Institut fr Instrumentelle Mathematik, Bonn, West Germany. See, also, Communication with Automata (in English). Griffiss Air Force Base, New York Technical Report RADC-Tr-65-377, 1, Suppl. 1. Petri, C.A. 1996. Nets, time and space, Theoretical Computer Science, Vol. 153, pp. 3-48. Peters, J.F. 1997. Time and clock information systems: Concepts and roughly fuzzy Petri net models. In: J. Kacprzyk (Ed.), Rough Sets and Knowledge Discovery. Berlin, Physica Verlag, a division of Springer Verlag [in press]. Peters, J.F.; Skowron, A.; Suraj, Z.; Pedrycz, W.; Ramanna, S. 1998. Approximate Real-Time Decision Making: Concepts and Rough Fuzzy Petri Net Models. International Journal of Intelligent Systems [to appear]. Peters, J.F.; Ziaei, K.; Ramanna, S. 1998. Approximate Time Rough Control: Concepts and Application to Satellite Attitude Control. In: Proc. Int. Conf. on Rough Sets and Current Trends in Computing (RSCTC98), Warsaw, Poland, June 1998 [to appear]. Skowron, A. 1995. Extracting laws from decision tables: a rough set approach. Computational Intelligence Vol. 11/2, pp. 371-388. Skowron, A.; Polkowski, L. 1996. Rough mereology: A new paradigm for approximate reasoning. Journal. of Approximate Reasoning Vol. 15/4, pp. 333-365. Skowron, A.; Rauszer, C. 1992. The discernability matrices and functions in information systems. In: Intelligent Decision Support, Handbook of Applications and Advances of the Rough Sets Theory, Slowinski, R. (Ed.), Dordrecht, Kluwer Academic Publishers (1992) 331-362. Skowron, A.; Suraj, Z. 1993a. A rough set approach to real-time state identification. Bulletin EATCS Vol. 50, pp. 264-275. Skowron, A.; Suraj, Z. 1993b. A rough set approach to real-time state identification for decision making. Institute of Computer Science Research Report 18/93, Warsaw University of Technology. Skowron, A.; Suraj, Z. 1994. Synthesis of concurrent systems specified by information systems. ICS Research Report 39/94, Institute of Computer Science, Warsaw Institute of Technology. Skowron, A.; Suraj, Z. 1995. Discovery of concurrent data models from experimental data tables: a rough set approach. Institute of Computer Science Research Report 15/95, Warsaw Institute of Technology. Skowron, A.; Suraj, Z. 1996. A parallel algorithm for real-time decision making: a rough set approach. Journal. of Intelligent Systems Vol. 7, pp. 5-28.

You might also like