1. NiSIS Malta Nov. 07 PERCEPTIVE SWARMS, DATA AND COMPUTABLE HABITATS: Swarm Intelligence in Clustering and Pattern Recognition Vitorino Ramos LaSEEB-IST, Evolutionary Systems and Biomedical Engineering Lab., Technical University of Lisbon, IST, Lisbon, PORTUGAL http://www.laseeb.org/vramos/ Perceptive Swarms, Data and Computable Habitats:

2. What is Swarm Intelligence? Swarm Intelligence (SI) is the property of a system whereby the collective behaviours of (unsophisticated) entities interacting locally with their environment cause coherent functional global patterns to emerge. SI provides a basis with which it is possible to explore collective (or distributed) problem solving without centralized control or the provision of a global model. To tackle the formation of a coherent social collective intelligence from individual behaviours, several bio-inspired concepts related to Self- Organization, Stigmergy and Social Foraging in animals are normally used. NiSIS Malta Nov. 07 Perceptive Swarms, Data and Computable Habitats:

3. Stigmergy: An example could be provided by two individuals, who interact indirectly when one of them modifies the environment and the other responds to the new environment at a later time. In other words, stigmergy could be defined as a typical case of environmental synergy. Grassé showed that the coordination and regulation of building activities do not depend on the workers themselves but are mainly achieved by the nest structure: a stimulating configuration triggers the response of a termite worker, transforming the configuration into another configuration that may trigger in turn another (possibly different) action performed by the same termite or any other worker in the colony. NiSIS Malta Nov. 07 Perceptive Swarms, Data and Computable Habitats:

4. A B La Valletta St. Julians A kind of Environmental Synergy A Collective geographic Memory Old Trails are used as memory, while the new ones are used for innovation, adaptation and for the construction of new feasible solutions. NiSIS Malta Nov. 07 Perceptive Swarms, Data and Computable Habitats:

5. Example - Swarms of social insects construct trails and networks of regular traffic via a process of pheromone (a chemical substance) laying and following. These patterns constitute what is known in brain science as a cognitive map. The main differences lies in the fact that insects write their spatial memories in the environment, while the mammalian cognitive map lies inside the brain. Forming a Cognitive Map NiSIS Malta Nov. 07 Perceptive Swarms, Data and Computable Habitats:

6. Natures trick: To combine signal reinforcement with its simultaneous evaporation NiSIS Malta Nov. 07 Perceptive Swarms, Data and Computable Habitats:

7. Memory and Robustness through Reinforcement Innovation through Evaporation NiSIS Malta Nov. 07 Perceptive Swarms, Data and Computable Habitats: Natures trick: To combine signal reinforcement with its simultaneous evaporation

8. 3. Trail forming model NiSIS Malta Nov. 07 Perceptive Swarms, Data and Computable Habitats:

9. Chialvo & Millonas model: - Is the simplest (local, memoryless, homogeneous and isotropic) model which leads to trail forming that we could find in the litterature, and the formation of trails and networks of ant traffic is not imposed by any special boundary conditions, lattice topology, or additional behavioral rules. - Its assumed that each organism emits pheromone at a given rate, and there is no spatial diffusion. Also global pheromone evaporates after all ants have moved at a given rate. - The ants are not allowed to have any memory and the individual’s spatial knowledge is restricted to local information about the pheromone density. - The pheromonal field (Cognitive map) contains information about past movements and decisions of the organisms, but not arbitrarily far in the past since the field “forgets” its distant history due to evaporation in time. - Toroidal boundary conditions are imposed on the lattice to remove, as far as possible, any boundary effects. - Nonlinear response or directional bias are introduced in order to form trails, or to persist on past trails that are already formed. 3. Trail forming model NiSIS Malta Nov. 07 Perceptive Swarms, Data and Computable Habitats:

10. Transition rule between cells by use of a pheromone weighting function: Measures the relative probabilities of moving to cell r with pheromone density, This parameter is associated with the osmotropotaxic sensitivity. Controls the degree of randomness with which the ant follows the gradient of pheromone. For low values the pheromone concentration does not greatly affect its choice, while high values cause it to follow pheromone gradient with more certainty. can be seen as the sensory capacity. This parameter describes the fact that the ant’s ability to sense pheromone at high concentrations. Chialvo & Millonas model: 3. Trail forming model Perceptive Swarms, Data and Computable Habitats:

11. Normalised Transition probabilities on the lattice to go from cell k to cell i: Measures the magnitude of the difference in orientation: w (0) = 1 w (1) = 1/2 w (2) = 1/4 w (3) = 1/12 w (4) = 1/20 Measures the relative probabilities of moving to cell i with pheromone density, e.g.: Coming from North w = 1/12 w = 1/4 w = 1/2 w = 1/20 w = 1 w = 1/4 w = 1/12 43 2 1 3 2 01 Indicates the sum over all the cells j which are in local neighbourhood of k. Chialvo & Millonas model: 3. Trail forming model NiSIS Malta Nov. 07 Perceptive Swarms, Data and Computable Habitats:

12. Coming from North w = 1/12 w = 1/4 w = 1/2 w = 1/20 w = 1 w = 1/4 w = 1/12 Coming from SouthWest w = 1 w = 1/2 w = 1/4w = 1/20 w = 1/2 w = 1/12 w = 1/4 Coming from NorthEast w = 1/20 w = 1/12 w = 1/4w = 1 w = 1/12 w = 1/2 w = 1/4 Coming from East w = 1/12 w = 1/20 w = 1/12w = 1/2 w = 1/4 w = 1 w = 1/2 E N S W Directional Bias (w) needed to compute Normalised Transition Probabilities on grey level 8 x 8 windows (some Examples): Chialvo & Millonas model: 3. Trail forming model NiSIS Malta Nov. 07 Perceptive Swarms, Data and Computable Habitats:

13. “Deciding where to go” by Roulette Wheel selection: w = 1 w = 1/2 w = 1/4w = 1/20 w = 1/2 w = 1/12 w = 1/4 w = 1/12 = 1 SW SSEWENWNNE Chialvo & Millonas model: 3. Trail forming model NiSIS Malta Nov. 07 Perceptive Swarms, Data and Computable Habitats:

14. Results: Parameters used: Pheromone deposition rate Pheromone evaporation rate Osmotropotaxic sensitivity Inverse of sensory capacity Chialvo & Millonas model: 3. Trail forming model NiSIS Malta Nov. 07 Perceptive Swarms, Data and Computable Habitats:

15. 4. Emerged Collective Perception One big CROSS or to many little SQUARES ?! Question: Can we solve this without a priori knowledge ?! NiSIS Malta Nov. 07 Perceptive Swarms, Data and Computable Habitats:

16. Extended model to grey level habitats -Instead of constant pheromone deposition rate, a term not constant is included: Pheromone deposition rate for a specific ant at a specific cell Chialvo & Millonas Pheromone deposition rate (constant)‏ Gives a measure of similarity between two different lattice windows, in terms of grey level spatial arrangement; 0 <= Dcm <= 1 (Matching Properties)‏ Constant 1 st Extension 4. Emerged Collective Perception NiSIS Malta Nov. 07 Perceptive Swarms, Data and Computable Habitats:

17. Extended model to grey level habitats -Pheromone deposition rate for a specific ant at a specific cell depends on grey level matching properties between 2 window lattices: 2 nd Extension Ant comes from Center and goes to NE: ?? 4. Emerged Collective Perception NiSIS Malta Nov. 07 Perceptive Swarms, Data and Computable Habitats:

18. Extended model to grey level habitats This measures differences on grey level overall intensity This measures differences on windows grey level homogeneity This measures successful matching properties between windows even considering all permutations; S equals the difference between the frequency of each class, for 2 grey level histograms (representing the 2 windows); Smax = 18. = 1 Max. variance difference is 126.711on 3x3 windows using 8 bit images Max. average difference is 255 using 8 bit images 4. Emerged Collective Perception NiSIS Malta Nov. 07 Perceptive Swarms, Data and Computable Habitats:

19. 4. Emerged Collective Perception NiSIS Malta Nov. 07 Perceptive Swarms, Data and Computable Habitats:

20. 4. Emerged Collective Perception Figure - Colony cognitive maps (pheromonal fields) for several iterations, on images Cross, Einstein, Map, Marble and Road. Except when indicated, parameters are those from [1]. In A2 and Z, ants are allowed to step on each other; habitats are respectively Cross and an homogeneous image. In this last case, results are similar with those found by Chialvo and Millonas [1]. A3) k=0.011. A4) k=0.019. A5)  =4.5. A6)  =2.5. NiSIS Malta Nov. 07 Perceptive Swarms, Data and Computable Habitats:

21. = ++ = ++ Ant_System_2D applied for 200 iterations (OUTPUT= Pheromone Distribution)‏ Negative Colour Image Channel Combining (Mono to RGB)‏ Channel Splitting (RGB to Mono)‏ RGB R G B

22. Original Colour Image Pheromone Distribution in Colour (after 200 iterations), the respective R,G and B channels, and the colour negative.

23. Adaptation and emerged perception between two images using self-organized swarms 4. Emerged Collective Perception NiSIS Malta Nov. 07 Perceptive Swarms, Data and Computable Habitats:

24. 4. Emerged Collective Perception Figure - One swarm (3000 ants) is thrown to explore Einstein image for 1000 iterations. At t=100, the Einstein habitat is replaced by Map image. Evolution of swarm cognitive maps (pheromonal fields) are shown for several iterations. NiSIS Malta Nov. 07 Perceptive Swarms, Data and Computable Habitats:

25. Fig. - Emerging pheromone maps in dynamic landscapes. The self- regulated swarm starts to evolve over "Einstein" image. After 100 iterations, the image changes to "Map". Adaptation and emerged perception between two images using self-organized swarms 4. Emerged Collective Perception NiSIS Malta Nov. 07 Perceptive Swarms, Data and Computable Habitats:

26. 5. MM Watershed vs. Swarm Image Perception NiSIS Malta Nov. 07 Perceptive Swarms, Data and Computable Habitats:

27. Fig. - Emerging pheromone maps in dynamic landscapes. The self-regulated swarm starts to evolve over "Einstein" image. After 100 iterations, the image changes to "Map". Emerging Perception and Adaptation between two different images WatershedWatershed+SVPSWatershed+SFPS WatershedWatershed+SVPSWatershed+SFPS 5 NiSIS Malta Nov. 07 Perceptive Swarms, Data and Computable Habitats:

28. 6. Swarm Intelligence based Gastric bypass Video Segmentation Figure. Typical film image: the stomach, the duodenum and the bypassed intestine are visible. NiSIS Malta Nov. 07 Perceptive Swarms, Data and Computable Habitats:

29. 6. Swarm Intelligence based Gastric bypass Video Segmentation Figure. Swarm Intelligence based (last column) versus Classical Mathematical Morphology Watershed based (second column) frame segmentation. 1st column) Original images (frame 100 up, and 220 down), 2nd column) Watershed segmentation, 3rd column) Pheromone distribution after applying the present proposal, and 4th column) final results after processing images on the 3rd column. NiSIS Malta Nov. 07 Perceptive Swarms, Data and Computable Habitats:

30. 7. Optimization In here, this additional term should naturally be related with specific characteristics of cells around one ant, like their altitude (z value or function value at coordinates x,y), having in mind our present aim. So, our pheromone deposition rate T, for a specific ant, at one specific cell i (at time t), should change to a dynamic value (p is a constant = 1.93) expressed by equation 3. In this equation, Δmax = | zmax – zmin |, being zmax the maximum altitude found by the colony so far on the function habitat, and zmin the lowest altitude. The other term Δ[i] is equivalent to (if our aim is to minimize any given landscape): Δ[i] = | zi – zmax |, being zi the current altitude of one ant at cell i. If on the contrary, our aim is to maximize any given dynamic landscape, then we should instead use Δ[i] = | zi – zmin |. NiSIS Malta Nov. 07 Perceptive Swarms, Data and Computable Habitats:

31. Application to Optimization Problems Ants are randomly placed on the landscape/fun ction. All ants move on each time step: the direction is chosen according to the pheromone levels around the ant and it is constrained by a directional bias. Environment is NxN toroidal grid with different values according to a function. Each time step, all ants deposit a certain amount of pheromone that is proportional to the value of the function on that site. t = 0t=1000t = 50t = 100t = 500 NiSIS Malta Nov. 07 Perceptive Swarms, Data and Computable Habitats:

32. 7. Optimization NiSIS Malta Nov. 07 Perceptive Swarms, Data and Computable Habitats:

33. 7. Optimization NiSIS Malta Nov. 07 Perceptive Swarms, Data and Computable Habitats:

34. 7. Optimization NiSIS Malta Nov. 07 Perceptive Swarms, Data and Computable Habitats:

35. Fig.- A self-organized swarm emerging a characteristic flocking migration behaviour between one deep valley (South region) and one peak (North region), surpassing in intermediate steps (Mickey Mouse shape) some local optima. Over each foraging step, the population self-regulates. 7. Optimization NiSIS Malta Nov. 07 Perceptive Swarms, Data and Computable Habitats:

36. Medium valleys Highest peak Medium valley Lowest valley Medium peak Medium peaks Medium valley Targets change: NiSIS Malta Nov. 07 Perceptive Swarms, Data and Computable Habitats:

37. 8. Bacterial foraging (BFOA) vs. Self-Regulated Swarms (SRS)‏ NiSIS Malta Nov. 07 Perceptive Swarms, Data and Computable Habitats:

38. 9. Dynamic Optimization When facing dynamic optimization problems the goal is no longer to find the extrema, but to track their progression through the space as closely as possible. Over these kind of over changing, complex and ubiquitous real-world problems, the explorative- exploitive subtle counterbalance character of our current state-of-the-art search algorithms should be biased towards an increased explorative behavior. While counterproductive in classic problems, the main and obvious reason of using it in severe dynamic problems is simple: while we engage ourselves in exploiting the extrema, the extrema moves elsewhere NiSIS Malta Nov. 07 Perceptive Swarms, Data and Computable Habitats:

39. 9. Dynamic Optimization NiSIS Malta Nov. 07 Perceptive Swarms, Data and Computable Habitats:

40. Figure - (LEFT) A 3D toroidal changing landscape describing a Dynamic Optimization (DO) Control Problem (8 frames in total). (RIGTH) A self- organized swarm emerging a characteristic flocking migration behaviour surpassing in intermediate steps some local optima over the 3D toroidal landscape above, describing a Dynamic Optimization (DO) Control Problem. Over each foraging step, the swarm self-regulates his population and keeps tracking the extrema (44 frames in total). 9. Dynamic Optimization NiSIS Malta Nov. 07 Perceptive Swarms, Data and Computable Habitats:

41. 9. Dynamic Optimization NiSIS Malta Nov. 07 Perceptive Swarms, Data and Computable Habitats:

42. 9. Dynamic Optimization

43. 9. Dynamic Optimization: Population Size NiSIS Malta Nov. 07 Perceptive Swarms, Data and Computable Habitats:

44. 9. Dynamic Optimization: Mean altitude NiSIS Malta Nov. 07 Perceptive Swarms, Data and Computable Habitats:

45. 11. Binary Ant Algorithm (BAA)‏ NiSIS Malta Nov. 07 Perceptive Swarms, Data and Computable Habitats:

46. 11. Binary Ant Algorithm (BAA)‏ NiSIS Malta Nov. 07 Perceptive Swarms, Data and Computable Habitats:

47. 12. Other Applications NiSIS Malta Nov. 07 Perceptive Swarms, Data and Computable Habitats:

48. Clustering and Classification 6 NiSIS Malta Nov. 07 Perceptive Swarms, Data and Computable Habitats:

49. Results within t = 1E6 time steps t = 1, E total = 2.910t = 50,000, E total = 1.264 t = 10,000, E total = 1.744t = 75,000, E total = 1.182 t = 20,000, E total = 1.513t = 1E6, E total = 0.906 FEATURES in 2D NiSIS Malta Nov. 07 Perceptive Swarms, Data and Computable Habitats: Clustering and Classification

50. t = 1E6, E total = 0.906 FINAL RESULT: NiSIS Malta Nov. 07 Perceptive Swarms, Data and Computable Habitats: Clustering and Classification

51. Results: NiSIS Malta Nov. 07 Perceptive Swarms, Data and Computable Habitats: Clustering and Classification

52. RESULTS: Latent semantic analysis and document filtering: preliminary results on newspapers (with JJ Merelo, 2002) The present work uses LSA as a feature extraction method, in order to map 931 words of an article at a Spanish newspaper. In the LSA model [6, 13], terms and documents are represented by an m x n incidence matrix A. Each of the n i unique terms in the document collection are assigned a row in the matrix, while each document is assigned a column. SVD is applied to the resulting matrix, and the main "axes" are them obtained. Words are projected onto those axes, resulting similar vector values for words with a similar meaning. Thus, each word uses a 50 feature vector. Since we had 931 items (words) to self-organize by the swarm, 91 ants were used, on a 61 x 61 non-parametric toroidal grid. Figure 5 shows the final result at t=10 6. Clustering for text Mining NiSIS Malta Nov. 07 Perceptive Swarms, Data and Computable Habitats:

53. (A) anunció, bilbao, embargo, titulos, entre, hacer, necesídad, tras, vida, lider, cualquier, derechos, medida.(B) dirigentes, prensa, ciu. (C) discos, amigos, grandes. (D) hechos, piloto, miedo, tipo, cd, informes. (E) dificil, gobierno, justicia, crisis, voluntad, creó, elección, horas, frente, técnica, unas, tarde, familia, sargento, necesídad, red, obra. (F) voz, puenlo, papel, asseguró. (G) nuestro, europea, china, ahora, poder, hasta, mucho, compañía, nacionalistas, cambio, asesinado, autor, nuevo, estamos, no. NiSIS Malta Nov. 07 Perceptive Swarms, Data and Computable Habitats: Clustering for text Mining

54. RESULTS: Image Retrieval Figure 5 – Spatial distribution of 244 images (representing 14 types of Portuguese Granites + 2 types of Chinese Granites), at t=1,000,000. Each image (point in the environment) is composed by 117 morphological and intensity features. Type 1 probability function was used with k 1 =0.1 and k 2 =0.3. Clustering for text Mining

55. DNA Protein Sequence Data Peng-Yeng Yin (Taiwan), Vitorino Ramos NiSIS Malta Nov. 07 Perceptive Swarms, Data and Computable Habitats:

56. Figure - Self-Organized Ant-based clustering results on IDS data (MIT Lincoln Labs) using a full data set with 11982 samples (41 features each) in the initial and final steps Intrusion detection Systems NiSIS Malta Nov. 07 Perceptive Swarms, Data and Computable Habitats:

57. Web Usage Mining Vitorino Ramos, Ajith Abraham (USA)‏ NiSIS Malta Nov. 07 Perceptive Swarms, Data and Computable Habitats:

58. Web Usage Mining Vitorino Ramos, Ajith Abraham (USA)‏ NiSIS Malta Nov. 07 Perceptive Swarms, Data and Computable Habitats:

59. Web Usage Mining Vitorino Ramos, Ajith Abraham (USA)‏ NiSIS Malta Nov. 07 Perceptive Swarms, Data and Computable Habitats:

60. Other Data Mining examples include: NiSIS Malta Nov. 07 - Ant-based Knowledge Discovery - ACO-based Data Mining - Construction of Rule-based Classifiers - Real Time Continuous Clustering - Ant-based Feature Selection and Extraction - Video Processing - Text and Document Mining - Web Usage Mining