Cooperative Interactive Cultural Algorithms Based on Dynamic Knowledge Alliance

Cooperative Interactive Cultural Algorithms Based on Dynamic Knowledge Alliance Yi-nan Guo 1, Shuguo Zhang 1, Jian Cheng 1,2, and Yong Lin 1 1 College of Information and Electronic Engineering, China University of Mining and Technology, Xuzhou, 221116 Jiangsu, China nanfly@126.com 2 Department of Automation, Tinghua University, Beijing 100084, China Abstract. In cooperative interactive genetic algorithms, each user evaluates all individuals in every generation through human-machine interface, which makes users tired. So population size and generation are limited. That means nobody can evaluate all individuals in search space, which leads to the deviation between the users best-liked individual and the optimal one by the evolution. In order to speed up the convergence, implicit knowledge denoting users preference is extracted and utized to induce the evolution. In the paper, users having simar preference are further divided into a group by K-means clustering method so as to share knowledge and exchange information each other. We call the group as knowledge alliance. The users included in a knowledge alliance vary dynamically whe their preferences are changed. Taken a fashion evolutionary design system as example, simulation results show that the algorithm speeds up the convergence and decreases the number of individuals evaluated by users. This can effectively alleviate users fatigue. Keywords: Cooperative Interactive Culture Algorithms, K-means, Dynamic Knowledge Alliance. 1 Introduction In interactive genetic algorithms (IGAs),the fitness values of individuals are evaluated by human[1]. The evolution operations of IGAs are usually implemented in a computer node for one user. Because human is easy to feel tired when he does one work for a long time, the generation and population size are limited. So it is difficult to obtain the satisfying solution in limited generation aiming at the optimization problems with large search space. Moreover, there are often more than one users participating in the evolution through human-machine interfaces in different computer nodes so as to obtain the satisfying individuals meeting themselves needs. In order to improve the performances of the evolution and effectively alleviate users fatigue, Guo[2] introduced dual structure of cultural algorithms into IGAs. Frequency pattern mining methods are adopted to extract key gene-meaning-unit which reflects users preferences in the evolution as implicit knowledge. Aiming at multiuser IGAs, many researchers proposed the methods to exchange the evolution information among users. This can improve the diversity of population in each clients and H. Deng et al. (Eds.): AICI 2011, Part III, LNAI 7004, pp. 204 211, 2011. Springer-Verlag Berlin Heidelberg 2011

Cooperative Interactive Cultural Algorithms Based on Dynamic Knowledge Alliance 205 the convergence. Based on the satisfying degree of group decision-making and users preferences, Sun [3] proposed a distributed collaborative interactive genetic algorithm. How to confirm the number of sharing individuals and select them are lustrated. Mitsunori [4] introduced asynchronous collaborative IGA. An elite sample database is but to solve the limits at time and space in parallel IGAs. Tomoyukii [5] gave the idea about collaboration into online shopping navigation system. The satisfying products are obtained automatically in terms of users current preferences. Based on interactive cultural algorithms (ICAs), Guo [6] further proposed cooperative ICAs. Each user s preference is noticed to other users by adopting IP multicast technology so as to exchange the evolution information in knowledge level. So the effective experiences of other users are used for reference in the co-evolution. In this method, the cost of network communication is increasing along with the increasing number of users. In order to avoid network congestion caused by knowledge sharing among all users, a clustering method for users based on knowledge alliance is proposed. The users with simar preferences are classified into a knowledge alliance. The information is noticed to all user inside the alliance by IP broadcast. Common knowledge reflecting all users common preferences is exchanged among alliances by tree network. This can effectively decrease the cost for information transfer. 2 Algorithm Description Multi-user cooperative interactive culture algorithm is simar to the traditional culture algorithms. They are essentially dual-layer evolution model. IGA is implemented in population space. In each independent computer node, human gives qualitative evaluation for all individuals via human-machine interface. The evolution operations including selection strategy, crossover operator and mutation operator are carried out based on the evaluated fitness values. Some better individuals are selected to belief space as samples by acceptance function in terms of certain rules. Obviously, these samples contain the information about users personal preference. In belief space, knowledge reflecting users preference is extracted from samples by frequency pattern mining in every independent computer node. It is shared with other users. Considering the communication cost of information migration among users, a hierarchical sharing mechanism of knowledge based on knowledge alliance is proposed by adopting construction method of agent alliances in multi-agent system. Fig. 1. Dynamic knowledge alliance in belief space

206 Y.-n. Guo et al. As shown in Fig.1, K-means clustering algorithm is used to partition users to knowledge alliances. Here, there are K alliances. Each user is described by a black dot in the alliance. Users belonging to the same knowledge alliance have simar preferences. The common knowledge is migrated among alliances in the manner of broadcast. When a alliance receives migrated information from other alliances, migrated knowledge wl be integrated with its own common knowledge. Then inconsistent information is noticed to all users in the alliance by IP broadcast. This hierarchical information exchange mechanisms can decrease network communication cost and effectively share evolution information among users. 3 Extraction and Utization of Knowledge in ICAs In ICAs, users preferences are extracted from the better individuals as knowledge. The individuals are usually encoded by binary. In existing knowledge extraction methods, one or more bits with large percentage are obtained by statistical methods directly. However, the relationship among the gene-meaning-units cannot be effectively reflected. So an extraction method based on frequency pattern mining is presented. Suppose [ 0, f max ] is the range of individuals fitness values. Let E (t) be the individual set meeting f ( x ) [ 0.5 f f i ]. The algorithm s steps are shown as follows: max, max Step1: In E (t), each individual is encoded by binary. Suppose there are G genemeaning-units in an individual. Each gene-meaning-units x ij is converted into nonoverlapping real coding x ij [8] and saved into database Q. x ij = Ω j + 1 + x, j = 0,1,2... G (1) ij Step2: After the database Q is scanned, the frequent items and their support counts is collected. Then FP-tree is constructed based on the minimum support counts. If FPtree contains a single path, the frequent patterns corresponding to the path are directly formed. Otherwise, the condition pattern database for each node is generated. The condition FP-tree is recursively obtained to get the frequent patterns. Step3: Each gene-meaning-units with certain values in obtained frequent patterns is transformed to corresponding gene-meaning-unit value with binary genotype by formula (2). Other uncertained bits are described as "*". x ij x ij Ω j 1, j = 0,1,2... = G (2) Here, the converted frequent patterns reflect the gene-meaning-units with largest frequency and maximum correlation in the evolution. They also note the most-liked favorite gene by users, called knowledge of users preferences. In this paper, frequent patterns are adopted to induce the evolution in next generation. If each bit of x ij is 0 or 1, the phenotype of this gene-meaning-unit is directly shown in human-machine interface to be evaluated. If x ij = * * is uncertain, it is add with own knowledge. If the uncertain genemeaning-unit stl exists, it is randomly generated by 0 or 1. In order to reserve own q

Cooperative Interactive Cultural Algorithms Based on Dynamic Knowledge Alliance 207 preference, the number of individuals influenced by migrated common knowledge are limited. Suppose knowledge influencing rate is γ ( 0Â1 ). Let pop is the population size. M is the number of individuals influenced by knowledge. The constrain for influenced population size is M γ pop. 4 Construction of Dynamic Knowledge Alliances Due to humans cognitive is variable, users preferences may change along with the evolution. So it s difficult to obtain the labeled training data reflecting user fixed preferences. Because K-means clustering method does not dependent on prior knowledge, a classification method for users by this algorithm is given. 4.1 Construction of Alliance Based on K-Means Clustering Method According to the knowledge extracted from computer nodes, the users having simar preference are collected by using K-means clustering method. A knowledge alliance for the users having common preference is constructed. Suppose { FP 1, FP2,, FP N } is users preferences set. N is the number of users. K users preferences are randomly selected as initial cluster centers expressed by C i ( i = 1,, K). The simarity of categories between the initial cluster center and jth users preference is calculated. S i j 1 = G K G δ ( FP jl, C ) (3) i = 1 l = 1 1 δ ( FP jl, C ) = 0 if FP if FP jl jl = C C (4) Here, δ ( FP jl, C ) describes the correlation between two frequent patterns. Based on the i simarity of categories, user is classified into the category with max S j. Thus, N users i are divided into K knowledge alliances according to their evolution knowledge. Then the cluster centers in all knowledge alliances denoted by Ci = C i1, Ci 2, C ig are adjusted according to the above classification results unt they are no longer changed. C = 1 N N i i i = 1 The number of users in a knowledge alliance is dynamically changed along with the evolution and the variation of users preferences. After the evaluation by users is finished in each generation, the simarity of categories between user s evolution knowledge and the corresponding cluster center is calculated. If the simarity of category is less than a certain threshold, the member of knowledge alliance does not change. If the simarity of category exceeds a certain threshold, this user exits from corresponding alliance. The simarity of categories between this user and other cluster centers are computed. The user joins an alliance with the maximum simarity. Then the cluster center of corresponding alliance is modified. C (5)

208 Y.-n. Guo et al. 4.2 Knowledge Migration Strategy In cooperative ICAs, hierarchical knowledge migration strategy is adopted. It contains the user-oriented knowledge migration within the alliance and the alliance-oriented knowledge migration among alliances. A) The migration mode After the evolution operations in each generation is completed, knowledge reflecting user s preference is extracted and sent to the server. The classification is done by the server before the evolution in next generation. Because human has own cognition and the degree of fatigue in the evaluation for individuals or the time spending on the evaluation is different, knowledge from different users may be asynchronous. So asynchronous knowledge migration is used in the paper. In asynchronous knowledge migration strategy, knowledge alliances may be constructed when knowledge from some of users is sent to the servers. But the maximum proportion of users participating in the classification must be ensured. Assume that T i is the evaluation time for ith user from the 1th generation to the generation sending knowledge to the sever at first time. If more than 0.8N users have submitted the knowledge to the sever, knowledge alliances are constructed for the first time. Users are gradually famiar with the system and know which kinds of individuals they like. So the evaluation time is decreasing. The interval for knowledge migration among alliances is defined by average T i, expressed by τ. Knowledge migration among alliances is done every iτ generation. N 1 = N i = 1 + 1 τ T i (6) B) Knowledge fusion method After an alliance receives common knowledge from other alliances, the knowledge is sent to all users in the alliance in the form of broadcast and used to induce the evolution. Here, the gene-meaning-unit is used as a unit. Migrated common knowledge and self knowledge are fused by and operation. Suppose g is the number of the genemeaning-units with certain values in self knowledge of alliance M i. g denotes the gene-meaning-units set with certain values from common knowledge of alliance M j. Then the effect of common knowledge on alliance M i is analyzed in two aspects: If g g, the self knowledge of the alliance is replaced by common knowledge from M j to induce the evolution. If g g and g g Φ, the genes with same values are retained. At the genes with different values, M i is replaced by the corresponding parts of the knowledge from M j. If g g and g g = Φ, common knowledge from M j has a direct impact on the evolution. c ) Communication complexity analysis of the algorithm In cooperative ICAs, a dual-layer communication structure is formed by the establishment of dynamic knowledge alliance. Compared with CICAs without knowledge alliance, the communication complexity is further analyzed as follows.

Cooperative Interactive Cultural Algorithms Based on Dynamic Knowledge Alliance 209 No classification for users: Suppose there are N users participating in the evolution. If users are not classified, the knowledge reflecting users preferences is notified to other users in the form of broadcast in each generation. Therefore, the communication complexity is N(N-1). Classification for users: Suppose there are K alliances formed in the evolution. If users are classified to knowledge alliances, the common knowledge reflecting users same preference in a alliance is notified to other users in the alliance in the form of broadcast. Two layer communication mode including information exchange within and among alliances is formed. So the communication complexity is N+K(K-1). Obviously, the larger N is, the communication complexity of CICA with knowledge alliances is less. This wl make the quantity of exchanged information less so as to effectively avoid network congestion in the communication process. 5 Simulation Results and Their Analysis In order to validate the rationality of the algorithm, fashion evolutionary design system is taken as examples. The goal of fashion evolutionary design system is to find the satisfying dress for users. A dress is composed of coat and skirt. Each part includes pattern and color, which are described by binary code with nine bits [12]. The same evolution operations and parameters are adopted in each computer node. Roulette selection method combing with elitist strategy are used. Multi-point crossover operator and one-point mutation operator are adopted. The detaed parameters values are: f max =100, pop=8, Pc=0.4, Pm=0.01, T=200, N=10, K=3. A) Dynamic analysis of knowledge alliance Ten users are participated in the evolution. The same evolutionary operations are done by everyone in each computer node. The simulation results are shown in Table.1. All users can find their own satisfying solutions. But the evaluation time and the evaluated individuals are different. That means everybody has different cognition. The knowledge alliance containing certain users varies along with the evolution, as shown in Table.2. Each user participates in the evolution asynchronously. The alliance containing them varies along with the evolution. If the evolution is finished in some computer node, corresponding user wl not belong to any alliance. Table 1. The simulation results User ID 1 2 3 4 5 6 7 8 9 10 iteration 5 11 12 12 9 6 10 6 8 9 evaluation time 4 58 9 19 9 10 6 52 5 54 5 08 7 33 6 07 6 36 7 55 evaluated individuals 36 68 76 61 53 45 76 44 58 67 whether the satisfying dress is found yes yes yes yes yes yes yes yes yes yes

210 Y.-n. Guo et al. Table 2. The variation of knowledge alliance times 1 2 3 4 1th alliance 4/5/8 1/5/8 6/9 4 2th alliance 1/2/10 2/3/4/10 3/10 3/10 3 th alliance 4/5/8 6/7/9 2/7/9 2/7 The variation of alliances cluster centers is shown in Table.3. Along with the evolution, one user may belong to different alliances. So the cluster centers may be changed. It is obvious that the number of certain gene-meaning-unit in cluster centers is increasing. That means the accuracy of common knowledge is increasing. Common knowledge is exchanged in an alliance or among the alliances so as to guide the evolution. This improves the diversity of population and accelerates the convergence. Table 3. The variation of cluster centers times 1th cluster center 2th cluster center 3th cluster center 1 **************1001 10111************* *****10001******** 2 01000*********1001 10111*****0011**** *****10001******** 3 01000*****01111001 10111*****0011**** 0010110001******** 4 01000*****01111001 10111*****00110111 00101100011010**** B) Comparison of performances among different algorithms In order to show the important role of knowledge alliance to improve the performances of cooperative ICAs(CICADKA), it is compared with cooperative interactive cultural algorithms without users clustering(cicas) and traditional interactive genetic algorithm(igas). The simulation results are shown in Table.4. Table 4. Comparison of performances among different algorithms algorithms generation time the evaluated individuals satisfaction rate CICADKAs 8.3 6 57 58.4 100% CIGAs 10.4 8 47 73 100% IGAs 15.8 9 52 99.1 70% CICADKA can effectively decrease the number of individuals evaluated by users and find satisfying solutions in a shorter time and smaller generation. Compared with traditional IGAs, CICADKA not only improves the rate of obtained satisfying individuals, but also accelerated the convergence at the same time. In conclusion, this algorithm is consistent with the principles of alleviating users fatigue.

Cooperative Interactive Cultural Algorithms Based on Dynamic Knowledge Alliance 211 6 Conclusions A multi-user cooperative ICAs with dynamic knowledge alliances is proposed. In each computer node, interactive cultural algorithms based on frequency pattern mining are done. Based on the knowledge given by nodes, users are dynamic classified to many alliances according to their preference by using K-means clustering method. Common knowledge is shared among these alliances. Knowledge is noticed to everyone inside the alliance by broadcast. Take fashion evolutionary design system as example, more than one user participates in the evolution. Simulation results show that the proposed algorithm can alleviate user s fatigue more and speed up the convergence. In a word, cooperative ICAs based on dynamic knowledge alliances is a feasible and effective multi-user interactive evolutionary algorithm. Acknowledgment. This work was supported by National Natural Science Foundation of China under Grant 60805025, Natural Science Foundation of Jiangsu under Grant BK2010183, the China Postdoctoral Science Foundation Funded Project under Grant 20090460328. References 1. Takagi, H.: Interactive evolutionary computation: fusion of the capabities of EC optimization and human evaluation. Proceedings of the IEEE 89(9), 1275 1296 (2001) 2. Guo, Y.-N., Lin, Y.: Interactive genetic algorithms with frequent-pattern mining. In: Proceedings of the 6th International Conference on Natural Computation, pp. 2381 2385 (2010) 3. Sun, X.-Y., Wang, X.-F., Gong, D.-W.: A distributed co-interactive genetic algorithm and its applications to group decision-making. Information and Control 36(5), 557 561 (2007) 4. Miki, M., Yamamoto, Y., Wake, S.: Global asynchronous distributed interactive genetic algorithm. In: 2006 IEEE International Conference on Systems, Man and Cybernetics, pp. 3481 3485 (2006) 5. Hiroyasu, T., Yokouchi, H.: Extraction of Design Variables using Collaborative Ftering for Interactive Genetic Algorithms. In: IEEE International Conference on Fuzzy Systems, pp. 1579 1584. IEEE, Piscataway (2009) 6. Guo, Y.-N., Lin, Y., Yang, M., Zhang, S.: User s preference aggregation based on parallel interactive genetic algorithms. Applied Mechanics and Materials Journal 34, 1159 1164 (2010) 7. Wang, J.P., Chen, H., Xu, Y., et al.: An architecture of agent-based intelligent control systems. In: Proceedings of the World Congress on Intelligent Control and Automation, pp. 404 407. IEEE, Piscataway (2000) 8. Le, M.N., Ong, Y.S.: A frequent pattern mining algorithm for understanding genetic algorithms. In: International Conference on Natural Computation, pp. 131 139. IEEE, Piscataway (2008)