Evolving a Disjunctive Predator Prey Swarm using PSO: Adapting Swarms with Swarms A thesis Presented by: Firasath Riyaz. Mentor: Dr. Peter M. Maurer. Co-Advisor: Dr. Robert J. Marks II 11/10/2005 Department Of Computer Science, BU, Waco.

Department Of Computer Science, BU, Waco. Outline Introduction Problem Solution Adapting Predator Prey Swarm Simulations, Results and Emergent Patterns Conclusions Possibilities for future work 11/10/2005 Department Of Computer Science, BU, Waco.

Department Of Computer Science, BU, Waco. Introduction Swarm Intelligence: “the emergent collective intelligence of groups of simple agents” Autonomous agents Simple rules Coordination without direct communication Unexpected complex group behavior 11/10/2005 Department Of Computer Science, BU, Waco.

Introduction (cont’d) 11/10/2005 Department Of Computer Science, BU, Waco.

Introduction (cont’d) Features: Distributed and multi-agent. Flexibility. Robustness. Self-organization Group behavior unexpected and complex Complex forms of social behavior can achieve a number of tasks. 11/10/2005 Department Of Computer Science, BU, Waco.

Introduction (cont’d) SI Applications Business Applications: Bonabeau, E. and Myer, C. (2001). Swarm Intelligence, A Whole New Way to think about Business, Harvard Business Review, pp. 106-114. Ant foraging for efficient routing of cargo, Southwest Airlines. Honey bees model for task allocation. “bucket-brigade”, enhanced efficiency and increased productivity Simple rules, self organization, Capital One 11/10/2005 Department Of Computer Science, BU, Waco.

Introduction (cont’d) Telecommunication: Bonabeau, E. and Myer, C. (2001). Swarm Intelligence, A Whole New Way to think about Business, Harvard Business Review, pp. 106-114. Ant foraging applied for efficient routing. Kassabalidis, I, El-Sharkawi, M.A., Marks, R.J. Asabshahi, P. and Gray, A.A.. (2001) Swarm Intelligence for Routing in Communication Networks, IEEE Globecom, San Antonio TX. AntNet algorithm for mobile ad-hoc networks. 11/10/2005 Department Of Computer Science, BU, Waco.

Introduction (cont’d) Optimization: Kennedy, J. and Eberhart, R. C. (1995). Particle swarm optimization, Proceedings of IEEE International Conference on Neural Networks, Piscataway, NJ. pp. 1942-1948. A function optimization algorithm Botee, H.M. and Bonabeau, E. ( 1999). Evolving ant colony optimization, Advances in Complex Systems, vol. 1, no. 2/3, pp. 149-159. Evolving ant colony optimization for solving TSP 11/10/2005 Department Of Computer Science, BU, Waco.

Introduction (cont’d) Robotics: Mondada F., Pettinaro G.C., Kwee I., Guignard A., Gambardella L.M., Floreano D., Nolfi S., Deneubourg J.-L. and Dorigo M. (2002) SWARM-BOT: A Swarm of Autonomous Mobile Robots with Self-Assembling Capabilities, Proceedings of the International Workshop on Self-Organisation and Evolution of Social Behaviour, pages 11-22, Monte Verità, Ascona, Switzerland Control a system governed by complex dynamics group of autonomous robots 11/10/2005 Department Of Computer Science, BU, Waco.

Problem Agent characteristics Autonomous Swarm based models have multiple agents Agent characteristics Autonomous Simple rules represented by sensors Robustness of the swarm lies in the autonomous or disjoint nature of agents. 11/10/2005 Department Of Computer Science, BU, Waco.

Department Of Computer Science, BU, Waco. Problem (cont’d) Important Issue in multi-agent systems Aggregation of simple rules in decision making Conjunctive aggregation Problem is if one sensor corresponding to a rule fails, the agent is ineffective 11/10/2005 Department Of Computer Science, BU, Waco.

Department Of Computer Science, BU, Waco. Solution Disjunctive aggregation: Since Added robustness: Linear degradation in agent’s performance under individual rules fail. Ability to evolve performance Emergent behavior Investigate robustness 11/10/2005 Department Of Computer Science, BU, Waco.

Department Of Computer Science, BU, Waco. Predator Prey Swarm Models a system governed by complex dynamics. Unpredictable effects of decision taken at agent level Explicit design of control is difficult Environment and agents A model to test Disjunctive aggregation. Predator models the behavior of a static environment Prey models the behavior of an agent 11/10/2005 Department Of Computer Science, BU, Waco.

Predator Prey Swarm (cont’d) World is defined by a two dimensional field Predator: Bully Current position, velocity, kill radius and neighborhood Follow a set of rules to capture prey. A predator does not die. Prey: Dweeb Current position, velocity and neighborhood. Follow a set of rules to evade the predators Maximize the median prey life 11/10/2005 Department Of Computer Science, BU, Waco.

Predator Prey Swarm (cont’d) Sample two dimensional field of 3 predators and 6 prey. 11/10/2005 Department Of Computer Science, BU, Waco.

Predator Prey Swarm (cont’d) Predator attack rules: Chase nearest prey Move towards Center of Mass of prey in neighborhood Random component to avoid coalescing 11/10/2005 Department Of Computer Science, BU, Waco.

Predator Prey Swarm (cont’d) Predator attack velocity attributes 11/10/2005 Department Of Computer Science, BU, Waco.

Predator Prey Swarm (cont’d) Predator velocity: Calculate unit velocity vectors in the three strategic attack directions. Each velocity vector has an associated weight Disjunctive aggregation of velocity components vnew = wnd * vnd + wdcom * vdcom + wr * vr Normalize vnew Physical constraint V = vold * momentum + vnew * ( 1 – momentum) 11/10/2005 Department Of Computer Science, BU, Waco.

Predator Prey Swarm (cont’d) Predator position update NewPosition = Old Position + MaxStepSize * V 11/10/2005 Department Of Computer Science, BU, Waco.

Predator Prey Swarm (cont’d) Prey survival attributes: move away from The nearest predator Center of mass of predators contained in the neighborhood. Move in a random direction to avoid coalescing/overlapping. Stay away from the boundaries to avoid being cornered. Move away from the nearest infected prey. The prey try to spread themselves in the field by staying away from the closest prey and moving away from center of mass prey in the neighborhood. 11/10/2005 Department Of Computer Science, BU, Waco.

Predator Prey Swarm (cont’d) Prey Survival Attributes 11/10/2005 Department Of Computer Science, BU, Waco.

Predator Prey Swarm (cont’d) Prey velocity: Calculate unit velocity vectors for different survival attributes. Fuzzy Linguistic Variables in a Fuzzy Inference Engine: Euclidean distance classified as VN, N, A, F, VF Each variable has a set of associated weights 11/10/2005 Department Of Computer Science, BU, Waco.

Predator Prey Swarm (cont’d) Disjunctive aggregation of prey velocity components vnew = wnb * vnb + wbcom * vbcom + wr * vr + wid * vid + wdcom * vdcom + wnd * vnd + wbt * vbt + wtp * vtp + wlw * vlw+ wrw * vrw 11/10/2005 Department Of Computer Science, BU, Waco.

Predator Prey Swarm (cont’d) Normalize vnew Physical Constraint. V = vold * momentum + vnew * ( 1 – momentum) Prey Position Update: NewPosition = Old Position + MaxStepSize * V A set of optimal weights for prey survival attributes are evolved using Particle Swarm Optimization 11/10/2005 Department Of Computer Science, BU, Waco.

Adapting Predator Prey Swarms Particle Swarm Optimization: Optimization algorithm imitating the motion of a flock of birds, or insects The system is initialized with a population of random solutions (particles) and searches for optima by updating generations Particles: position and velocity. Fitness 11/10/2005 Department Of Computer Science, BU, Waco.

Adapting Predator Prey Swarm (cont’d) PSO Algorithm: For each particle     Initialize particle END Do     For each particle         Calculate fitness value         If the fitness value is better than the best fitness value (pBest) in history        set current value as the new pBest     End     Choose the particle with the best fitness value of all the particles as the gBest     For each particle         Calculate particle velocity according to equation 1         Update particle position according to equation 2     End While maximum iterations or minimum error criteria is not attained. 11/10/2005 Department Of Computer Science, BU, Waco.

Adapting Predator Prey Swarm (cont’d) vnew = vold + c1 * r1 * (pbest - present) + c2 * r2 * (gbest - present) (1) present = present + vnew (2) Where: vnew and vold is the new and old velocities of the particle, present: Current position of the particle in search space. pbest: particle position which gave the best fitness value so far. gbest: best of pbests so far. r1 and r2 are uniform random numbers in [0, 1]. c1 and c2 are learning factors usually set to 2. 11/10/2005 Department Of Computer Science, BU, Waco.

Adapting Predator Prey Swarm (cont’d) PSO was used to evolve prey survival attribute weights. Each particle represents a set of weights For each particle predator prey swarm was repeated several times Fitness function optimized: Average median prey life 11/10/2005 Department Of Computer Science, BU, Waco.

Simulations, Results and Emergent Patterns Simulation setup Parameters. Results Emergent patterns 11/10/2005 Department Of Computer Science, BU, Waco.

Simulations, Results and Emergent Patterns Simulation setup Software configuration: Java Hardware configuration: Dell Pentium 4, 1.8 Ghz, 512MB machines in the graduate student lab, Computer Science Dept. 11/10/2005 Department Of Computer Science, BU, Waco.

Simulations, Results and Emergent Patterns Parameters used Predator prey swarm Parameter Description Value Field Size The size of the field in which predator prey swarm is evolved. 400 x 400 (pixel applet window) Number of predators The number of predators in the swarm. 10 Number of prey The number of prey in the swarm. 100 11/10/2005 Department Of Computer Science, BU, Waco.

Simulations, Results and Emergent Patterns Predator parameters Prey parameters Parameter Value Kill-radius 5 ( pixel positions) Com-radius 200 (pixels) Max step size Momentum 0.75 Parameter Value Infected prey distance 100 ( pixel positions) Com-radius 30 (pixels) Max step size 3 ( pixel positions) Momentum 0.25 11/10/2005 Department Of Computer Science, BU, Waco.

Simulations, Results and Emergent Patterns PSO parameters Fitness function optimized: Average of median prey life. Parameter Description Value Number of particles Number of the particles in PSO. 10 Vmax Maximum particle velocity 1 c1 and c2 Learning factors 2 Predator-Prey repeat For each particle number of times predator prey swarm is evolved in one PSO generation. 5 11/10/2005 Department Of Computer Science, BU, Waco.

Simulations, Results and Emergent Patterns PSO Simulation 1: 234 generations The Global Best and Average Personal Best Fitness function is plotted 11/10/2005 Department Of Computer Science, BU, Waco.

Simulations, Results and Emergent Patterns Evolved set of prey survival attribute weights Parameter VN N A F VF wnb 1.496112 -0.26885 0.326199 0.139085 3.104299 wbcom -0.19884 3.022029 0.780692 0.422761 -0.52131 wdcom 2.978585 -0.74418 0.731056 0.767199 -1.32521 wnd 0.383726 0.033212 -0.5205 1.176138 -0.63978 wlw -2.32374 2.182005 -0.1704 0.368014 -0.3302 wrw 0.924111 0.15717 1.167848 0.725385 -1.55305 wtp 0.366553 0.983607 -1.48285 -0.70846 -0.68069 wbt -0.0634 2.114983 -1.7776 -1.09796 -2.37636 wid -0.50897 -0.75386 -1.94931 0.831417 0.706812 11/10/2005 Department Of Computer Science, BU, Waco.

Simulations, Results and Emergent Patterns PSO Simulation 2: 122 generations The Global Best and Average Personal Best Fitness function is plotted 11/10/2005 Department Of Computer Science, BU, Waco.

Simulations, Results and Emergent Patterns Evolved set of prey survival attribute weights Parameter VN N A F VF wnb 2.362566 -1.55364 1.543917 -1.46267 -3.00413 wbcom 0.396071 0.283025 -0.33709 1.573843 -3.35227 wdcom 2.857888 1.300653 0.157554 1.64632 -0.26407 wnd 0.085772 0.634661 -0.9144 1.645024 -0.94478 wlw 0.969846 -2.66092 -2.7327 0.534569 -1.45258 wrw 0.052089 2.2854 0.907741 0.135652 -0.73609 wtp 1.01798 0.574888 -2.57348 -1.69388 -1.45157 wbt 0.760509 0.798146 1.113263 0.472713 -0.56423 wid -0.99737 -0.14158 -2.91363 0.536843 1.148609 11/10/2005 Department Of Computer Science, BU, Waco.

Simulations, Results and Emergent Patterns PSO Simulation 3: 112 generations The Global Best and Average Personal Best Fitness function is plotted 11/10/2005 Department Of Computer Science, BU, Waco.

Simulations, Results and Emergent Patterns Evolved set of prey survival attribute weights Parameter VN N A F VF wnb 0.321632 -1.07872 0.781259 -0.16421 1.880079 wbcom 1.18288 -0.09487 -0.243 -0.36856 1.466024 wdcom 4.451869 3.083868 -1.91994 0.732869 1.374435 wnd 1.069739 -2.07995 -0.85324 0.50287 0.041671 wlw 1.541971 0.692149 0.802267 -0.00504 -0.93835 wrw 0.773392 -1.77489 -3.93243 0.361954 0.589858 wtp 1.851252 0.807819 1.953532 1.634385 -1.16497 wbt -1.40294 0.308706 -1.45974 0.751738 -0.14512 wid 0.449467 2.145414 0.278686 1.605107 -0.15784 11/10/2005 Department Of Computer Science, BU, Waco.

Simulations, Results and Emergent Patterns Simulation 2 produced better global fitness value Consistency check: 3 different predator prey simulations Predator-prey swarm repeat: 25 times Simulation 1 gives best optimal fitness Weights from simulation 3 are more consistent Weights used from Average of median prey fitness Standard Deviation Simulation one 3496134 3442460 Simulation two 853655.8 2728819 Simulation three 1762688 763701 11/10/2005 Department Of Computer Science, BU, Waco.

Simulations, Results and Emergent Patterns The predator prey swarm was simulated with the most consistent optimal set of weights. Unexpected behavior emerged. Self sacrificing prey Prey evolved negative weights Reversal of rule Increased swarm aggregate life 11/10/2005 Department Of Computer Science, BU, Waco.

Simulations, Results and Emergent Patterns Self sacrificing prey Self sacrificing prey 11/10/2005 Department Of Computer Science, BU, Waco.

Simulations, Results and Emergent Patterns Prey herding predators Prey evolved a strategy to surround the predators Vacillating predator motion and herding Increased swarm aggregate life 11/10/2005 Department Of Computer Science, BU, Waco.

Simulations, Results and Emergent Patterns Prey herding predators 11/10/2005 Department Of Computer Science, BU, Waco.

Simulations, Results and Emergent Patterns Prey circular motion Prey move in circles Cannot be cornered Increased swarm aggregate life 11/10/2005 Department Of Computer Science, BU, Waco.

Simulations, Results and Emergent Patterns Prey circular motion 11/10/2005 Department Of Computer Science, BU, Waco.

Simulations, Results and Emergent Patterns Robustness of the evolved swarm is tested Varying parameters Number of predators 5 to 100 in steps of 5 Number of prey 25 to 80 in steps of 5 11/10/2005 Department Of Computer Science, BU, Waco.

Simulations, Results and Emergent Patterns Gradual degradation 11/10/2005 Department Of Computer Science, BU, Waco.

Simulations, Results and Emergent Patterns Robustness of disjunctive aggregation in rule failures is studied Cumulative dropping of rules Different predator prey population sizes Number of swarm simulations 10 Predator step size=10, prey step size = 6 11/10/2005 Department Of Computer Science, BU, Waco.

Simulations, Results and Emergent Patterns 11/10/2005 Department Of Computer Science, BU, Waco.

Simulations, Results and Emergent Patterns There is a gradual decrease in fitness of swarm aggregate There is no sudden collapse of the swarm Robustness Disjoint agents Disjoint rules 11/10/2005 Department Of Computer Science, BU, Waco.

Simulations, Results and Emergent Patterns 11/10/2005 Department Of Computer Science, BU, Waco.

Simulations, Results and Emergent Patterns Analysis of single prey against one or multiple predators Robustness only due to disjoint aggregation Agent does not collapse under rule failures 11/10/2005 Department Of Computer Science, BU, Waco.

Department Of Computer Science, BU, Waco. Conclusions Disjunctive Aggregation Disjoint at two levels Two fold robustness Robust under agent failure Robust under rule failure Ability to evolve Emergent patterns 11/10/2005 Department Of Computer Science, BU, Waco.

Possibilities for future work Further investigation into robustness of evolved swarm for varying parameter values Field size, step size etc. Incorporation of weak conjunctive logic between a subset of rules 11/10/2005 Department Of Computer Science, BU, Waco.

Department Of Computer Science, BU, Waco. “with all human achievements in science and engineering, nature still provides the best systems that can ever be fashioned” – A philosopher Questions 11/10/2005 Department Of Computer Science, BU, Waco.