International Journal of Machine Learning and Computing, Vol. 4, No. 5, October 2014 Symbiotic Particle Swarm Optimization for Neural Fuzzy Controllers Cheng-Hung Chen and Wen-Hsien Chen the population. The SPSO in which each particle represents a Abstract—This study proposes a symbiotic particle swarm single fuzzy rule differs from original symbiotic evolution [8] optimization (SPSO) for the specific neural fuzzy controller to adopt a multiple swarm scheme. A fuzzy system with (NFC). The specific NFC model using compensatory fuzzy R-rules is constructed by selecting and combining R particles operators of neural fuzzy networks makes fuzzy logic systems more adaptive and effective. The proposed SPSO adopts a from each swarm, and allowing the rule itself to evolve. multiple swarm scheme that uses each particle represents a single fuzzy rule and each particle in each swarm evolves separately to avoid falling in a local optimal solution. II. STRUCTURE OF NEURAL FUZZY CONTROLLER Furthermore, the SPSO embeds the symbiotic evolution scheme In this section, the structure of the specific NFC is in a specific particle swarm optimization (PSO) to accelerate the search and increase global search capacity. introduced. Compensatory operators in the specific NFC model are used to optimize fuzzy logic reasoning and to select Index Terms—Water bath temperature system, neural fuzzy optimal fuzzy operators. Therefore, an effective NFC should networks, symbiotic evolution, particle swarm optimization. not only adaptively adjust its fuzzy membership functions, it should also dynamically optimize adaptive fuzzy operators. y I. INTRODUCTION Neural fuzzy controllers (NFC) [1], [2] have been demonstrated to solving many engineering problems. They Layer 5 combine the capability of neural networks to learn from (Output nodes) processes and the capability of fuzzy reasoning under linguistic information pertaining to numerical variables. On Layer 4 … the other hand, recent development in genetic algorithms (Consequent (GAs) has provided a method for neural fuzzy network design. nodes) N N N w01 wi1xi w02 wi2 xi w03 wi3 xi Genetic fuzzy systems (GFSs) [3], [4] hybridize the i1 i1 i1 approximate reasoning of fuzzy systems with the learning Layer 3 C C C … capability of genetic algorithms. Furthermore, a new (Compensatory rule nodes) optimization algorithm, called particle swarm optimization (PSO), appears to be better than the genetic algorithm. It is an evolutionary computation technique that was developed by Layer 2 … … (Membership Kennedy and Eberhart in 1995 [5]. The underlying motivation function nodes) for the development of PSO algorithm is the social behavior of animals, such as bird flocking, fish schooling and swarm theory. PSO has been successfully applied to many Layer 1 (Input nodes) optimization problems, such as control problems [6], [7]. This study proposes a symbiotic particle swarm x1 x2 optimization (SPSO) for the specific neural fuzzy controller Fig. 1. Structure of the specific NFC model. (NFC). The specific NFC is based on our previous research [8] with adaptive compensatory fuzzy reasoning to dynamically Fig. 1 shows the structure of the specific NFC model. The adjust fuzzy operators. The proposed SPSO embeds the NFC realizes a fuzzy rule in the following form: symbiotic evolution scheme in a particle swarm optimization j (PSO) to accelerate the search and increase global search 1 j Rule j : IF[x is A and x is A and x is A and x is A ] N capacity. Unlike the GAs in a population as a full solution to a 1 1 j 2 2 j i ij N Nj THEN y' is w w x w x w x w x problem, symbiotic evolution [9] assumes that each 0 j 1 j 1 2 j 2 ij i Nj N individual in a population represents only a partial solution to (1) a problem. Complex solutions combine several individuals in where xi is the input variable; y' is the input variable; Aij is the linguistic term of the precondition part; j [0,1] is the Manuscript received April 30, 2014; revised May 22, 2014 C. H. Chen and W. H. Chen are with the Department of Electrical compensatory degree, w0 j and wij are the corresponding Engineering, National Formosa University, Yunlin County, Taiwan, ROC weight of feedback in the consequent part. (e-mail: [email protected]). DOI: 10.7763/IJMLC.2014.V4.450 433 International Journal of Machine Learning and Computing, Vol. 4, No. 5, October 2014 YMBIOTIC ARTICLE WARM PTIMIZATION FOR THE III. A S P S O where yd represents the dth model output; yd represents the NFC MODEL desired output, and D represents the number of input data. This section describes the proposed SPSO for the specific Step 3: Divide the fitness value by R and accumulate the NFC model. The SPSO method comprises of three major divided fitness value to the fitness record of the R selected components – create initial swarms, evaluate fitness value and rules with their recorded fitness values initially set to zero. update each particle. First, the initial swarms are created Step 4: Repeat the above steps until each rule (particle) in randomly before the evolution process begins. Second, the each swarm has been selected a sufficient number of times, fitness value is evaluated for each particle by symbiotic and record the number of fuzzy systems to which each particle evolution strategy to allow the rule itself to evolve in each has contributed. swarm. Third, the each particle is updated using local best, Step 5: Divide the accumulated fitness of each particle by global best and cooperative best in SPSO. the number of times it has been selected. A. Create Initial Swarms Particle m m ….. m ….. j w w w ….. w ….. 1j 1 j 2j 2 j ij ij 0j 1j 2j ij 1) Coding step The foremost step in SPSO is the coding of a fuzzy rule into a particle. Fig. 2 shows an example of the coding of Rule 11 Rule 12 parameters of a fuzzy rule into a particle where i and j . represent the ith input variable and the jth rule, respectively. Swarm 1 . Rule 1k . In this study, a Gaussian membership function is adopted with . variables that represent the mean and deviation of the Rule 1ps . membership function. Fig. 2 represents a fuzzy rule given by . Rule 1 …... Rule j …... Rule R NFC 1 . Rule j1 . Eq. (1), where mij and ij are the mean and variance of a . Rule j2 . Gaussian membership function, respectively; j is the . Swarm j . compensatory factor, and w is the corresponding parameter Rule jk . ij . of the consequent part associated with the jth rule node. In this . Rule jps study, a real number represents the position of each particle. Rule 1 …... Rule j …... Rule R NFC Nr . Particle . Rule R1 Rule R m1j m2j ….. mij ….. w0j w1j w2j ….. wij ….. 2 1 j 2 j ij j . Fig. 2. Coding a fuzzy rule into a particle. Swarm R . Rule R k . Cbest 1 …... Cbest j …... Cbest R 2) Initial swarms . Rule Rps Cooperative Best (Cbest) Before the SPSO method is applied, every position x j,k must be created randomly in the range [0, 1] in each swarm, Fig. 3. Structure of particle in the proposed SPSO method. where j=1, 2, …, R represents the jth swarm and k=1, 2, …, ps C. Update Each Particle represents the kth particle. Step 1: Update local best Lj,k, global best Gj, and B. Evaluate Fitness Value cooperative best Cj This subsection presents a novel method of symbiotic The local best position Lj,k is the best previous position that evolution. As described above, in the symbiotic evolution, the yielded the best fitness value of the jth swarm of the kth fitness value of a rule (a particle) is computed as the sum of particle, the global best position Gj is generated by the whole the fitness values of all the feasible combinations of that rule local best position and the cooperative best position Cj is with all other randomly selected rules, and then dividing this created by parameters of the best composed fuzzy system. In sum by the total number of combinations. Fig. 3 shows the step 1, the first step updates the local best position. Compare structure of the particle in the symbiotic evolution. In this the fitness value of each current particle with that of its local figure, the best parameters of fuzzy system are reserved (i.e., best position. If the fitness value of the current particle the best particle) by the cooperative best (Cbest). The exceeds those of its local best position, then the local best stepwise assignment of the fitness value is as follows. position is replaced with the position of the current particle. Step 0: Divide the rules into swarms of size ps. The second step updates the global best position. Compare the Step 1: Randomly select R fuzzy rules (particles) from each fitness value of all particles in their local best positions with of the above swarms, and compose the fuzzy system using that of the particle in the global best position. If fitness value these R rules. of the particle in the local best position is better than those of Step 2: Calculate fitness value of the composed fuzzy the particles in the global best position, then the global best system. In this study, the fitness value is given by the follow position is replaced with the current local best position. formula; x j,k , if F (x j,k ) F (L j,k ) 1 L j,k Fitness value (2) L , if F (x ) F (L ) 2 j,k j,k j,k 1 D . (3) 1 y y d d G j arg max F(L j,k ), 1 k ps L D d 1 j ,k 434 International Journal of Machine Learning and Computing, Vol.