Presentation on theme: "Evolution: Games, dynamics and algorithms"— Presentation transcript:
1Evolution: Games, dynamics and algorithms Karen PageBioinformatics UnitDept. of Computer Science, UCL
2EvolutionDarwinian evolution is based on three fundamental principles: reproduction, mutation and selectionConcepts like fitness and natural selection are best defined in terms of mathematical equationsWe show how many of the existing frameworks for the mathematical description of evolution may be derived from a single unifying framework
3Summary of what will be discussed Games, evolutionary game theoryKey frameworks of evolutionary dynamicsDeriving a unifying frameworkAn application to Fisher’s Fundamental TheoremRelationship with Genetic Algorithms
4What is game theory?Formal way to analyse interactions between agents who behave strategicallyMathematics of decision making in conflict situationsUsual to assume players are “rational”Widely applied to the study of economics, warfare, politics, animal behaviour, sociology, business, ecology and evolutionary biologyThe objective of each player is to maximise his/her score in the game. Maximising one players score may well not be consitent with maximsing anothers.Classical game theory looks at how a perfectly “rational” players will play the game. Rational means that the players try to maximize their scores in the game.
5Assumptions of game theory The game consists of an interaction between two or more playersEach player can decide between two or more well-defined strategiesFor each set of specified choices, each player gets a given score (payoff)Expand on specified choices: player 1 plays strategy a, player 2 plays strategy b etc.
6The Prisoners’ Dilemma Probably most studied of all gamesNot enough evidence to convict two suspects of armed robbery, enough for theft of getaway carBoth confess (4 years each), both stay quiet (2 years each), one tells (0 years) the other doesn’t (5 years)Stay quiet= cooperate (C) ; confess = defect (D)Payoff to player 1:Police have arrested two people whom they know have a committed an armed robbery, but they don’t have enough evidence for a jury to convict them. They do however, have evidence that the people stole a getaway car. They offer the suspects a deal. If neither of them confesses to the armed robbery they both get convicted of the theft of the car and do 2 years. If both of them confesses then they both share the blame do 4 years for armed robbery. If one confesses and the other doesn’t then the tell-tale gets set free and his mate does 5 years for armed robbery. So what should a prisoner do?R is REWARD for mutual cooperation =3S SUCKER’s payoff =0T TEMPTATION to defect =5P PUNISHMENT for mutual defection=1with T>R>P>S
7The problem of cooperation What ever player 2 does, player 1does better by defecting:Classical game theory both players DShame because they’d do better by both cooperatingCooperation is a very general problem in biologyEveryone benefits from being in cooperative group, but each can do better by exploiting cooperative efforts of others
8Trade wars and cartels Import tariffs - Should countries remove them? Price fixing- why not cheat?As mentionned at the beginning, game theory is applied to many disciplines. The prisoners dilemma occurs in many walks of life. Here I mention just a couple of examples.
9Repeated gamesIn many situations, typically players interact repeatedly- repeated Prisoners DilemmaStrategies can involve memory, use reciprocityTit-for-tatPavlovWhat one player does in round 4 of the game can depend on what has happened in the previous three rounds.
10Game theory and a computer tournament Game theory says it is rational to defect in single game or fixed number of roundsAxelrod’s tournament- double victory for Tit-for-TatFirst point is a repeat of what was said in “The problem of cooperation”Backwards induction
11Evolutionary Game Theory So how can cooperation be explained?
12Evolutionary games John Maynard Smith- evolution of animal behaviour Behaviour shaped by trial and error- adaptation through natural selection or individual learningPlayers no longer have to be ‘rational’: follow instincts, procedures, habits rather than computing best strategy.Games played in a population. Scores are summed. Strategies which do well against the population on average propagate.Phenotypic approach to evolutionFrequency-dependent selectionAt one time it was thought that rational behaviour would prove optimal against “irrational” behaviour. This turned out not to be the case.Ecology: A scarcity of prey will cause predators to starve and their numbers to decline. Having fewer predators around will favour the growth of the population of prey which in turn will allow more predators to survive. But now higher predator numbers will cause the number of prey to decrease, bringing us full circle. Hence in ecological systems, numbers of predators and prey can oscillate in time, because of the effects that the levels of each have on the other.
13Simple evolutionary game simulations Everyone starts with a random strategyEveryone population plays game against everyone elseThe payoffs are added upThe total payoff determines the number of offspring (Selection)Offspring inherit approximately the strategy of their parents (Mutation)[Note similarity to genetic algorithms.][Nash equilibrium in a population setting- no other strategy can invade]Talk about darwin’s theory of evolution- the frequencies of genes and increase over time if they are associated with features which lead to the production of more offspring. So the proportion of the given feature within the population will increase over time, eg. finches with a gene for sharp beaks in an environment where such a peak is very important for accessing food.If the feature is behavioural, such as for instance the propensity to back down in a conflict (cf. hawk-dove game) then whether the gene is selected for depends on the make up of the population. Evolutionary game theory models this kind off evolutionary process. It can be used to show that the proportion of hawks in a population of hawks and doves will tend to fitness gain for winning territory/ fitness loss for getting injured (in our example 1/5). Here it implies that heavily in armed species, such as stags, which can potentially inflict mortal wounds on one another, very few individuals will escalate a conflict. Paradoxically in species of doves who under normal circumstances can’t do each other much damage, escalation is much more likely. Indeed when confined to small cages doves will often peck each other to death.
14Evolution in the Prisoners’ Dilemma Standard evolutionary game (random interactions) all DefectModifications- spatial games: Interactions no longer random, but with spatial neighbours:Sum scores. Player with highest score of 9 shaded takes square (territory, food, mates) in next generationSome degree of cooperation evolves!Mention also proportional selection
15Simulations of the spatial Prisoners Dilemma 75 generationsWinner-takes-all selectionNo mutationRed=d(d last) Blue=c(c last) Yellow=d(c last) Green=c(d last)
16Conclusions on Evolutionary Games Game theory can be applied to studying animal and human behaviour (economics - evolutionary biology).Often traditional game theory’s assumption of ‘rationality’ fails to describe human/ animal behaviourInstead of working out the optimal strategy, assume that strategies are shaped by trial and error by a process of natural selection or learning. This can be modelled by evolutionary game theory.Space can matter
19The replicator equation Replicator equation describes evolution of frequencies of phenotypes within a population with fitness-proportionate selectionEg. game theory, replicators like “Game of Life”Frequency of type i is and fitness of type i is then
20The equivalence with Lotka Volterra equations Lotka Volterra systems of ecology describe the numbers of animals (eg. fish) of different species and are of the form:where is the abundance of species i, its fitness and there are n species in total.Often these interacting species oscillate in abundance.There is a precise equivalence with the replicator system for (n+1) types given by the substitution
21Replicator equation with mutation and quasispecies Suppose there are errors in replicating. The probability of type j mutating to type i isWe obtain a replicator equation with mutation:The equivalent with numbers rather than frequencies of types isWhen the fitnesses do not depend on frequencies, this is the quasispecies eqn. (Probably the case in most GAs?)
22Quasispecies equation Describes molecular evolution (Eigen)N biochemical sequencesBiochemical species i has frequency yiReplication at rate fi is error-prone - mutation to type j at rate qij
23Adaptive dynamics framework Game consists of a continuous space of strategies (eg.)Population is assumed to be homogeneous- all players adopt same strategyMutation generates variant strategies very close to the resident strategyIf a mutant beats the resident players it takes over otherwise it is rejectedAdaptive dynamics illustrates the nature of evolutionary stable strategies
24Adaptive dynamics equations Strategies are described by continuous parameters :Expected score of mutant against S is given by E(S’,S)The adaptive dynamics flow in the direction which maximises the score:
25We can derive Price’s equation from replicator-mutator equation Price’s equation from population genetics describes any type of selection.Suppose an individual of type i, frequency , has some trait p of value, so using the replicator equation with mutation we obtainThis applies when the values of are const.[p is the expected mutational change in p.]
27Price’s equation gives rise to adaptive dynamics If we assume that the mutation is localised and symmetrical then we can neglect the second term in Price’s eqn.Assume population is almost homogeneous and fitness is differentiable then we can Taylor expand the fitness, obtainingcf. adaptive dynamics:
29Fisher’s fundamental theorem Suppose fitnesses of genotypes constant. Can consider f as the trait p and obtain (for symmetric mutation):Fisher’s fundamental theorem of NSIn general, fitnesses of genotypes depend on environment. In game theory context, depend on the frequencies of other genotypes. Fisher’s theorem doesn’t apply- eg. PD
30Generalized version where We can use Price’s equation to obtain a generalized version of Fisher’s fundamental theorem:whereThis applies when the s depend linearly on the frequencies of genotypes- normally the case in evolutionary game theory.
31Fisher’s theorem and GAs In most GAs, fitnesses of particular solutions (chromosomes) probably fixed and so (except for the complication of recombination) Fisher’s theorem should hold:So for a GA with fitness-proportionate selection, no recombination and fixed fitness for a given solution, the average fitness of the population of solutions increases until there is no diversity left in the fitnesses.
32Conclusions on unifying evolutionary dynamics Unifying frameworkDifferent frameworks for different problems.We derive from Price’s equation a generalized version of Fisher’s Fundamental Theorem of Natural Selection.The Price – replicator framework can also be applied to discrete time formulations and to formulations with sexual reproduction.
34Evolutionary games and genetic algorithms Two-way interaction:1) So far discussed computer simulations of evolutionary processes, eg. evolution of animal behaviour2) Evolutionary computation, eg. genetic algorithms = computer science based on theory of biological evolutionEvolutionary games very like genetic algorithms- but1) Population size is usually quite large and may be few phenotypes: space well searched but not v. efficient.2) Usually no recombination3) Fitnesses depend on interactionsRefer to Mark Herbster’s course. Ask students about course.
35Genetic AlgorithmsEvolutionary models are computer algorithms which use evolutionary methods of optimisation to solve practical problems (cf. finding stable strategies in games rather than working out ‘rational’ solution)- eg. Evolutionary programming, genetic algorithmsEvolutionary operations involved in genetic algorithms: selection, mutation, recombination:Explain selection and recombination
36How evolutionary dynamics relates to GAs GAs evolve by selection and mutation their dynamics can be (to some extent) described by the replicator equation with mutation (cf. unifying framework).The replicator equation describes fitness-proportionate selection.Ficici, Melnik and Pollack (2000) - effects of different types of selection (eg. truncation) on the dynamics of the Hawk-Dove game + relevance for evolutionary algorithms. Can lead to different dynamics.Must also consider the effects of recombination.Ficici, S.G., Melnik, O., and Pollack, J.B. (2000) "A Game-Theoretic Investigation of Selection Methods Used in Evolutionary Algorithms." In Proceedings of the 2000 Congress on Evolutionary Computation. Zalzala, A., et al (eds.). IEEE Press.
37Incorporating recombination into the replicator framework Do this by assuming that rjk;i = probability that when parent chromosome of type j combines with parent chromosome of type k, an offspring of type i is formed.No mutation, recombination after replication:[NB discrete-time version]
38Adding in mutationAdd in mutation. Assume, as before, is probability type i mutates to form type j ( large). Assume this happens after recombination.What we had before wasWhat we have now is
39The diversity of the population and adaptive dynamics From Fisher’s theorem, see that no diversity of fitness in population no further increase in average fitness.However, because the variation in the parameters of the your system has become very small (population convergence), does not mean no further evolution.In the case of small variation, we can apply the adaptive dynamics framework which shows how the average values of traits (parameters) will change in time
40Relationship: evolutionary games & GAs - Conclusions Often evolution leads in the long run to ‘optimal’ solutions, like Nash equilibria.Ability of evolutionary processes to seek out optimal strategies has been exploited in computer science by the development of genetic algorithms and evolutionary computation for problem solving.Comparing with the use of computer simulations to study biological evolution, we see that there is a two-way interaction between biological evolutionary theory and computer science.
41Relationship to GAs- Conclusions Frameworks of evolutionary dynamics can be applied to GAs by modifying them to include recombination.Which framework is most informative depends on the individual problem, but we have shown they are equivalent.Eg. can look at detailed dynamics using the replicator-mutator frameworkOr we can look at a “converged” population using the adaptive dynamics framework.Looking further at the relationship between GAs and evolutionary dynamics could yield new solutions/ techniques for both.
42Acknowledgements Martin Nowak (IAS, Princeton) Terry Leaves (BNP Paribas, London)Karl Sigmund (Univ. Vienna)Steven Frank (Univ. California, Irvine)Peter Bentley (UCL)Christoph Hauert (Univ. British Columbia)Anargyros Sarafopoulos (Univ. Bournemouth)Bernard Buxton (UCL)To do: Look at Lande - quantative genetics stuff - G covariance matrixStochastic dynamicsCoevolutionary dynamics