XFUZZY 3.0: A development environment for fuzzy systems

XFUZZY 3.0: A DEVELOPMENT ENVIRONMENT FOR FUZZY SYSTEMS F. J. Moreno Velo I. Baturone S. Sánchez Solano A. Barriga Instituto de Microelectrónica de Sevilla - Centro Nacional de Microelectrónica Avda. Reina Mercedes s/n, (Edif. CICA) E-41012, Sevilla, Spain International Conference in Fuzzy Logic and Technology (EUSFLAT 2001), pp. 93-96, Leicester, Sep. 2001 This material is presented to ensure timely dissemination of scholarly and technical work. Copyright and all rights therein are retained by authors or by other copyright holders. All persons copying this information are expected to adhere to the terms and constraints invoked by each author’s copyright. In most cases, these works may not be reposted without the explicit permission of the copyright holder. Abstract This paper presents the new version of Xfuzzy, Xfuzzy 3.0, which is a development environment for fuzzy-inference-based systems. It is composed by many tools that cover the different stages of the fuzzy system design process, from their initial description to the final implementation. Its main features are the capability for developing complex systems, the flexibility of allowing the user to extend the set of available functions and the possibility of been executed on any platform with JRE (Java Runtime Environment) installed. Keywords: CAD Tools, Fuzzy System Development, Fuzzy System Tuning, Fuzzy Circuits. 1. Introduction Fuzzy systems have become an important alternative in many fields of application, thus increasing the necessity of easing their development [4][5]. In the last few years several CAD tools tailored to the fuzzy system design have been created. Nevertheless, many of these tools are dedicated to specific realizations and have constraints on the set of fuzzy operations they support, the complexity of the systems they can develop and their capability for automatic tuning, simulation or synthesis. Our research group has been working these years on the development of general purpose CAD tools which could solve these limitations. This led us to create Xfuzzy 2.0 [1], an environment which integrates different tools for edition, tuning, simulation and synthesis of fuzzy systems. These tools share a formal specification language, called XFL [2], which deals with the definition of complex systems. The new version of the environment attempts to dive into this line of work. For this purpose, new tools covering distinct aspects of the design process have been created and a more homogeneous graphical interface has been defined. Moreover, the specification language has been extended. The new language, called XFL3 [3], incorporates linguistic hedges and allows extending the set of available fuzzy functions. The environment has been completely programmed in Java, so it can be executed on any equipment with JRE (Java Runtime Environment) installed. 2. The Xfuzzy 3.0 environment Xfuzzy 3.0 is a development environment for fuzzy systems which integrates several tools covering the different stages of the design (Fig. 1). The description stage includes graphical tools for the definition of systems. The tuning stage consists in applying learning algorithms. The verification stage is composed of tools for simulation, monitorization and graphical representations of the system behavior. Finally, the synthesis stage is divided into tools generating high-level languages descriptions for software or hardware implementations. The nexus between all these tools is the use of a common specification language, XFL3, which extends the capabilities of XFL, the language defined in version XFUZZY 3.0: A Development Environment for Fuzzy Systems F. J. Moreno Velo I. Baturone S. Sánchez Solano A. Barriga Instituto de Microelectrónica de Sevilla Centro Nacional de Microelectrónica - CSIC Avda. Reina Mercedes, s/n. Edif. CICA, E-41012, Sevilla, Spain [email protected] XFL3 Specification Xfedit System editor Xfpkg Package editor Xfc C synthesis Description stage Xfcpp C++ synthesis Xfj Java synthesis Xfvhdl VHDL synthesis Synthesis stage Xf2dplot Graphical representation Xf3dplot Graphical representation Xfmt Inference monitor Xfsim Simulation Verification stage Xfsl Supervised learning Xfrl Reinforcement learning Tuning stage Figure 1: Design flow using Xfuzzy 3.0. (Shaded rectangles refer to tools under development) This work has been partially supported by the Spanish CICYT Project TIC98-0869 and the FEDER Project Nº 1FD97-0956-C3-02. 2.0. XFL3 is a flexible and powerful language, which allows to express very complex relations between the fuzzy variables, by means of hierarchical rule bases and user-defined fuzzy connectives, linguistic hedges, membership functions and defuzzification methods. Every tool can be executed as an independent program. The environment integrates all of them under a graphical user interface which eases the design process (Fig. 2). 3. Description stage The first step in the development of a fuzzy system is to select a preliminary description of the system. This description will be refined lately as a result of the tuning and verification stages. Xfuzzy 3.0 contains two tools assisting in the description of fuzzy systems: xfedit and xfpkg. The first one is dedicated to the logical definition of the system, that is, the definition of its linguistic variables and the logical relations between them. On the other side, the xfpkg tool eases the description of the mathematical functions assigned to the fuzzy operators, linguistic hedges, membership functions and defuzzification methods. The description of a fuzzy system in XFL3 language consists of several objects defining operator sets, linguistic variable types and rule bases, as well as a description of its hierarchical structure. The xfedit tool offers a graphical interface to ease the description of systems, avoiding the need for a deep knowledge of the XFL3 language. The system hierarchy is shown on the main window of xfedit (Fig. 3). From this window, other edition windows for operator sets, variable types and rule bases can be displayed. The edition of an operator set is carried out by selecting the mathematical functions assigned to the distinct logical connectives, rule implication and aggregation operators, linguistic hedges and defuzzification methods (Fig. 4). The definition of a linguistic variable type includes the description of its universe of discourse, of the labels applied to the variable and the membership functions assigned to those labels. The type edition window (Fig. 5) allows the introduction of the minimum/maximum values and the cardinality, which defines respectively the limits and the number of elements of the universe of discourse. Moreover, the window eases the definition of the linguistic labels and their membership functions, showing graphically the distribution of these functions on the universe of discourse and giving facilities for modifying their parameters. The rule bases contain the expressions of the logical relations between the linguistic variables of the system. The XFL3 language can describe very complex relations between the variables by combining basic propo- Figure 2: Main window of Xfuzzy 3.0. Figure 3: Main window of xfedit Figure 4: Operatorset edition on xfedit. Figure 5: Type edition on xfedit. sitions (which compares the value of a variable with a linguistic label) via connectives and linguistic hedges. The window for editing rule bases on xfedit (Fig. 6) presents three formats for the rule definition: matrix form, table form and free form. The matrix form constricts the rule base to two inputs and one output, and shows the rules in a compact form as a matrix. Each matrix element represents a rule like «if a is A & b is B then c is C». The table form is valid for any number of inputs and outputs, and each element of the table represents a rule like «if x0 is X0 & x1 is X1 & ... & xn is XN then z is Z». Finally, the free form allows exploiting the whole power of XFL3 to define complex relations like «if x0 is greater than X0 or not x3 is strongly equal to X3 then z is Z». One of the main features of the Xfuzzy 3.0 environment is the possibility of extending the set of mathematical functions that can be assigned to the different fuzzy connectives, linguistic hedges, rule aggregation, implication operators, available membership functions, and defuzzification methods. These mathematical functions are defined in files named as packages. The xfpkg tool is dedicated to the edition of function packages, which can be used by all the environment tools (Fig. 7). The definition of a new mathematical function requires the introduction of the information needed by any tool of the environment. This includes the definition of the function parameters and their constraints, the description of the behavior in the high-level languages Java, C y C++ (needed by the software synthesis) and the description of the differential function (required by some learning algorithms of the tuning stage). 4. Tuning stage Tuning of fuzzy system behavior usually becomes is always one of the most complex tasks of the design process. This behavior depends on the logical structure of the rule bases and on the definition of the membership functions employed for the linguistic variable types. Apart from unforeseen redesigns, the tuning process often deals with the modification of the parameters defining the membership functions. Since this should be performed simultaneously over a great number of parameters, a manual procedure happens to be an extremely complex task, making it necessary the use of automated tuning techniques. Xfuzzy 3.0 environment currently contains one tool dedicated to the tuning stage: xfsl, which is based on the use of supervised learning algorithms (Fig. 8). Another tool dedicated to reinforcement learning procedures is under development. Both tools intend the system behavior to approximate a known behavior. In supervised learning, a set of desired input/output patterns is available, which defines an error function between the system behavior and the desired one. Supervised learning attempts to minimize this error function. When resorting to the use of reinforcement learning, the exact values of system outputs are unknown. Reinforcement learning is based on studying the on-line behavior of the system to discover the effect the system must produce on its operational environment. The xfsl tool includes gradient descent algorithms, conjugate gradient algorithms, quasi-Newton algorithms, descent without derivatives, and stochastic Figure 6: Rulebase edition on xfedit. Figure 7: Main window of xfpkg. Figure 8: Main window of xfsl. methods like simulated annealing. The tool admits different error functions and allows the selection of the parameters to be tuned. Besides, xfsl includes two methods of pre- and post-processing to simplify the fuzzy system. The first one consists on eliminating non significant rules and labels. The second one develop a clustering over the labels of the output variables. 5. Verification stage The objective of the verification stage is to study the behavior of the system under development, detecting probable deviations on the expected behavior and identifying the source of these deviations. Xfuzzy 3.0 environment currently contains two verification tools: xf2dplot and xf3dplot, which are dedicated to graphically represent the system behavior in two (xf2dplot) or three dimensions (xf3dplot). Both tools work in a similar way: the input and output variables to be represented must be selected while values of the non-represented variables must be assigned (Fig. 9). Other tools under development are xfmt and xfsim. The former attempts to monitorize the system, showing the activation degree of every linguistic label and logical rule, as well as the value of the different inner variables, for some determined input values. The xfsim tool is aimed at simulating the system within its actual or modeled operational environment. It allows illustrating the system evolution by means of graphical representations of user-selected variables. 6. Synthesis stage The final stage of a fuzzy system design is to generate a system implementation, either software or hardware. Selecting the proper solution depends on the complexity of the system and the imposed requirements. When inference speed, area and power requirements are not restrictive, software implementations ar preferred to offer a great flexibility and the possibility of incorporating the system into a more general software project. Xfuzzy 3.0 environment provides the user with three tools for software synthesis: xfc,xfcpp,xfj, which generate descriptions of the fuzzy systems in C, C++ and Java high-level languages, respectively. On the other side, when system requirements are restrictive, hardware implementations are more adequate. To aid the designer in this task, another tool which is currently under development is xfvhdl. The purpose of this tool is to generate an VHDL description of the system, able to be synthesized as an ASIC (Application Specific Integrated Circuit) or an FPGA (Field Programmable Gate Array). 7. Conclusions The version 3.0 of Xfuzzy presented in this paper is a complete environment for the design of fuzzy systems. Its tools (some of them under development) covers efficiently the different stages of the design process. These tools share a common formal description language which allows the definition of complex systems by means of hierarchical rule bases and user-config- urable membership functions, defuzzification methods, fuzzy connectives, linguistic hedges and implication and aggregation operators. The environment is distributed freely under the GNU General Public License and, since it has been programmed entirely in Java, it can be executed on any platform containing the Java Runtime Environment (JRE). References [1] López, D.R., Jiménez, C.J., Baturone, I., Barriga, A., Sánchez-Solano, S.. "Xfuzzy: A Design Environment for Fuzzy Systems", Proc. 7th IEEE Int. Conf. on Fuzzy Systems (FUZZIEEE’98), pp. 1060-1065, Anchorage, May 1998. [2] López, D.R., Moreno, F.J., Barriga, A., Sánchez-Sol- ano, S.. "XFL: A Language for the Definition of Fuzzy Systems", Proc. 6th IEEE Int. Conf. on Fuzzy Systems (FUZZIEEE’97), pp. 1585-1591, Barcelona, Jul., 1997 [3] Moreno, F.J., Sánchez-Solano, S., Barriga, A., Baturone, I., López, D.R., “XFL3: Un nuevo lenguaje de especificación de sistemas difusos”, Actas X Congreso Español sobre Tecnologías y Lógica Fuzzy (ESTYLF’2000), pp.509-514, Sevilla, 2000. [4] Munakata, T., Jani, Y., Fuzzy systems: an overview, Communications of the ACM, Vol. 37, N. 3, 1994. [5] Sugeno, M., Ed., Industrial Application of Fuzzy Control, North-Holland, pp. 19-40, 1985. Figure 9: Main window of xf3dplot.