Competitive Co-evolution

Similar documents
Co-evolving predator and prey robots: Do arms races arise in artificial evolution?

NEUROEVOLUTION. Contents. Evolutionary Computation. Neuroevolution. Types of neuro-evolution algorithms

Co-evolutionary dynamics on a deformable landscape

1 Co-Evolution and Ontogenetic Change in Competing Robots

COMPETITIVE CO-EVOLUTION OF SENSORY-MOTOR SYSTEMS

Continual Coevolution through Complexification

Competitive Coevolution through Evolutionary Complexification

Evolutionary Design I

Toward Open-Ended Evolutionary Robotics: Evolving Elementary Robotic Units able to Self- Assemble and Self-Reproduce

Coevolution of Predators and Prey in a Spatial Model

Embodied Evolution and Learning: The Neglected Timing of Maturation

Swarm-bots. Marco Dorigo FNRS Research Director IRIDIA Université Libre de Bruxelles

IV. Evolutionary Computing. Read Flake, ch. 20. Assumptions. Genetic Algorithms. Fitness-Biased Selection. Outline of Simplified GA

SimAnt Simulation Using NEAT

V. Evolutionary Computing. Read Flake, ch. 20. Assumptions. Genetic Algorithms. Fitness-Biased Selection. Outline of Simplified GA

V. Evolutionary Computing. Read Flake, ch. 20. Genetic Algorithms. Part 5A: Genetic Algorithms 4/10/17. A. Genetic Algorithms

A comparison of evolutionary and coevolutionary search

Exercise 3 Exploring Fitness and Population Change under Selection

Normalised evolutionary activity statistics and the need for phenotypic evidence

Evolutionary computation

Analyzing Cooperative Coevolution with Evolutionary Game Theory

Evolutionary computation

Evolution of Adaptive Synapses: Robots with Fast Adaptive Behavior in New Environments

Swarm-bots and Swarmanoid: Two experiments in embodied swarm intelligence

Structure Learning of a Behavior Network for Context Dependent Adaptability

Complexity in social dynamics : from the. micro to the macro. Lecture 4. Franco Bagnoli. Lecture 4. Namur 7-18/4/2008

w1 w2 w3 Figure 1: Terminal Cell Example systems. Through a development process that transforms a single undivided cell (the gamete) into a body consi

Chapter 8: Introduction to Evolutionary Computation

Neural Systems and Artificial Life Group, Institute of Psychology, National Research Council, Rome. Evolving Modular Architectures for Neural Networks

Evolutionary Computation. DEIS-Cesena Alma Mater Studiorum Università di Bologna Cesena (Italia)

Complex Systems Theory and Evolution

Forming Neural Networks Through Efficient and Adaptive Coevolution

Composable Group Behaviors

Assessing Three-Dimensional Learning in the Next Generation Science Standards

Coevolution of Multiagent Systems using NEAT

Sexual Reproduction. Page by: OpenStax

How Species Form. 4.3 How Species Form. Reproductive Isolation

Introduction to Digital Evolution Handout Answers

A comparison of evolutionary and coevolutionary search

Repeated Occurrences of the Baldwin Effect Can Guide Evolution on Rugged Fitness Landscapes

Chapter 7. Evolutionary Game Theory

EVOLUTIONARY STABILITY FOR TWO-STAGE HAWK-DOVE GAMES

A Simple Approach to the Multi-Predator Multi-Prey Pursuit Domain

Co-evolution of Morphology and Control for Roombots

Homeotic Genes and Body Patterns

Chapter 1 Introduction

EVOLUTIONARY GAMES AND LOCAL DYNAMICS

Sexual Reproduction *

The Origin of New Species

Evolutionary Dynamics

Topics. Antibiotic resistance, changing environment LITERACY MATHEMATICS. Traits, variation, population MATHEMATICS

The Effect of Connection Cost on Modularity in Evolved Neural Networks

These are my slides and notes introducing the Red Queen Game to the National Association of Biology Teachers meeting in Denver in 2016.

Super Selection. 7 th Grade. Concepts. Objectives. Outline. Duration Pre-Visit: 40 minutes Museum Visit: 60 minutes Post Visit: 50 minutes

Transfer of Neuroevolved Controllers in Unstable Domains

biologically-inspired computing lecture 18

Neuroevolution for sound event detection in real life audio: A pilot study

5 LIFE HOW DO WE DEFINE LIFE AND WHAT DO WE KNOW ABOUT IT?

COMP3411: Artificial Intelligence 7a. Evolutionary Computation

Quantitative Description of Robot-Environment Interaction Using Chaos Theory 1

[Read Chapter 9] [Exercises 9.1, 9.2, 9.3, 9.4]

LINEAGE ACTIVITIES Draft Descriptions December 10, Whale Evolution

Finding Attack Strategies for Predator Swarms Using Genetic Algorithms

CS 570: Machine Learning Seminar. Fall 2016

Artificial Ecosystems for Creative Discovery

Extending the Temporal Horizon of Autonomous Robots

Theory of Evolution. Chapter 15

ESRM 350 Evolution: a brief review

Balancing and Control of a Freely-Swinging Pendulum Using a Model-Free Reinforcement Learning Algorithm

Evolutionary Bargaining Strategies

Maintenance of Species Diversity by Predation in the Tierra System

Chapter 2 Section 1 discussed the effect of the environment on the phenotype of individuals light, population ratio, type of soil, temperature )

From soup to cells the origin of life

EVOLVING PREDATOR CONTROL PROGRAMS FOR A HEXAPOD ROBOT PURSUING A PREY

A.I.: Beyond Classical Search

Harnessing Evolution: The Interaction Between Sexual Recombination and Coevolution in Digital Organisms

Computational Ecology Introduction to Ecological Science. Sonny Bleicher Ph.D.

Evolutionary Game Theory

An Analysis of Diploidy and Dominance in Genetic Algorithms

Information Entropy Theory of Physics

Chapter 15 Evolution

Enhancing a Model-Free Adaptive Controller through Evolutionary Computation

Improving Coordination via Emergent Communication in Cooperative Multiagent Systems: A Genetic Network Programming Approach

Environmental signals

Parallel/Distributed Evolutionary Computation The influence of spatial interaction in evolutionary behaviour

Experimental Evolution in Avida-ED Michigan State University NSF Beacon Center for the Study of Evolution in Action

Sympatric Speciation in Framsticks Virtual Life Research Project. Anske van Luijtelaar Sjoerd Kranedonk David Smits Joachim de Greeff

Learning Communication for Multi-agent Systems

Coevolution of learning and niche construction. Reiji Suzuki, Yasunori Noba and Takaya Arita

Predator escape: an ecologically realistic scenario for the evolutionary origins of multicellularity. Student handout

EVOLUTION change in populations over time

Proceedings of Neural, Parallel, and Scientific Computations 4 (2010) xx-xx PHASE OSCILLATOR NETWORK WITH PIECEWISE-LINEAR DYNAMICS

RL 3: Reinforcement Learning

Haploid & diploid recombination and their evolutionary impact

Development. biologically-inspired computing. lecture 16. Informatics luis rocha x x x. Syntactic Operations. biologically Inspired computing

Evolving gene regulatory networks controlling foraging strategies of prey and predators in an artificial ecosystem

EVOLUTION change in populations over time

COMP3411: Artificial Intelligence 10a. Evolutionary Computation

Evolution Notes Darwin and His Ideas

Forecasting & Futurism

Transcription:

Competitive Co-evolution Robert Lowe Motivations for studying Competitive co-evolution Evolve interesting behavioural strategies - fitness relative to opponent (zero sum game) Observe dynamics of evolving populations (evolutionary stability, e.g. evolutionary game theory). Motivations for studying Competitive co-evolution Behavioural Strategies: increasing complexity in behavioural strategies greater variety of strategies Dynamics of co-evolving populations: oscillatory dynamics helps to avoid local minima may be a powerful means for discovering optimized solutions stable strategies vs stable ratios (Maynard Smith, 1982 Evolution and the Theory of Games ) 1

Co-evolution for discovery of interesting behavioural strategies Karl Sims Creatures (1994): Evolution of competitive strategies regarding occupation of a block. Co-evolution of behavioural strategies Sims (1994) All contests are 1 on 1. A number of different ways of pairing competitors of different populations are used: see figure 2 (from Sims (1994)) Co-evolution of behavioural strategies - Sims (1994) 2

Dynamics of Competitive coevolution - The Red Queen Effect The Red Queen was a living chess piece in Lewis Carroll s Through the looking glass, who ran perpetually without getting very far because the landscape kept up with her (Cliff & Miller 1995) Fitness co-dependence and Arms races competitive co-evolution (CCE) ~ the evolution of two or more competing populations with coupled fitness e.g. predator - prey may enhance the power of artificial evolution evolutionary arms races ~ competing populations can drive each other to increasing levels of behavioral complexity a pedagogical series of challenges that gradually require more complex solutions hypothesized by biologists to be one of the main sources of evolutionary innovation and adaptation BUT CCE can create trivial oscillations between simple behaviours Incremental Evolution Vs CCE Incremental Evolution is a gradual modification of the fitness function or the environment but: requires careful attention and planning by the experimenter Natural (and possibly artificial) CCE, on the other hand, The environment contains another network that is also evolving So the task effectively faced by each species can become increasingly more complex Throughout evolution, agents face opponents that use different strategies Therefore they must themselves develop different counterstrategies or more general abilities 3

The Red Queen Difficult to measure/assess progress of the co-evolutionary process any solution/strategy found in one generation could be no longer valid in later ones the fitness landscape changes all the time (!) difficult to monitor progress with conventional indicators such as best and average fitness values Monitoring CCE Techniques that have been suggested CIAO (Current Individual vs. Ancestral Opponents) (Cliff & Miller, 1995) - testing the performance of the best (elite) individual of each generation against the best competing ancestors from each previous generation an Evolutionarily Stable Strategy (ESS)? Master tournament (F & N, 1997) - testing against all best opponents, even future ones (i.e. only possible after the evolutionary process) Hall of fame tournament (Rosin & Belew 1997) adopt elitism to encourage arms races. Predator-Prey Experiments 2 Khepera robots predator has vision prey is twice as fast 64 pixel linear scanner with 36 o field of view. Reduced to a 5 bit visual field 4

Predator-Prey Experiments Fitness Measures Fitness graphs show oscillatory trends, but do not inform us of: Evolutionary progress An appropriate optimization approach Frequency dependency strategies can account for oscillatory patterns Predators discover and rediscover two classes of strategy (all replications): A1 - track the prey and try to catch it A2 - track the prey while remaining in one area and attacking only on promising occasions Prey cycles between: B1 - stay still or hide near a wall; try to escape when detecting the predator B2 - move fast, avoiding both predator and walls A1 > B1 (late detection); B2 > A1 (faster prey); A2 > B2 (waiting for too fast prey); B1 > A2 (let the predator wait ) 5

Monitoring Co-Evolution In these experiments agents repeatedly switched between two opposing strategy sets and evolution did not really progress but It is claimed that co-evolution should be able to enhance artificial evolution by adding incrementally through arms races promoting generality Changing landscapes reduce the local minima problem 6

Master Tournament Like CAIO but elites compete against future elites too. (filling in the CAIO square) This allows us to plot the actual fitness relative to past and future agents Hall of Fame Every individual of this generation competes against the elite of previous generations to ascertain their fitness for reproduction Master Tournament plot for the Individuals generated by a hall of fame selection strategy CCE vs Individual Population Evolution: Evolving One Species Only 7

Co-Evolution and Learning Evolutionary Adaptation: No learning. Ontogenetic Adaptation: Genotype encodes sign, learning rate, learning rule but weights are always initialized to small random values in all runs predators show higher average and best fitness values Co-Evolution and Learning plastic predators adapt their strategies during their lifetime (e.g. center) almost all of them can adapt to B1 and B2 the prey usually do not (or cannot because of sensorimotor limitations) because they have less time and less information for learning when learning was optional predators evolved to learn, prey did not (!) Summary Advantages of Competitive Co-Evolution (CCE): allows the study of adaptation in a changing environment, this sometimes promotes a greater diversity of behavioural strategies. under some conditions, may produce non-supervised incremental evolution (when a general solution exists). Disadvantage of CCE: May find limit cycles. Possible solution to cycling problem: Elitism but may limit evolutionary possibilities constrained by evolutionary history, biologically unrealistic CCE can drive towards general strategies exploited further by learning. Caveat: Constraints imposed by design or by body sensor morphology have a strong effect on the behaviours evolved 8

Group Discussion We want to evolve general strategies that work well in a variety of situations. This may require search and refinement of agent strategies rather than have the agents constantly switch between different strategies How can we promote this? What other factors might aid co-evolution? How would you extend this work? (think about nature) References Nolfi & Floreano (2000) Evolutionary Robotics. Cambridge, MA: MIT Press. - chapter 8 Cliff D. &Miller G.F (1995) Tracking the Red Queen: Measurements of adaptive progress in co-evolutionary simulations. In F. Moran, A. Moreno, JJ. Merelo, and P. Chacon (Eds.) Advances in Artificial Life: Proceedings of the third european conference on artificial life. Berlin: Springer Verlag. Floreano & Mondada (1998). Evolutionary neurocontrollers for autonomous mobile robots. Neural Networks, 11(7-8). Predator/Prey videos: http://gral.ip.rm.cnr.it/nolfi/predprey/ K.Sims, Artificial Life IV Proceedings, ed.by Brooks & Maes, MIT Press, 1994, pp.28-39. K.Sims video: http://www.archive.org/details/sims_evolved_virtual_creatures_1994 9