An Introduction to Differential Evolution. Kelly Fleetwood

Transcription

1 An Introduction to Differential Evolution Kelly Fleetwood

2 Synopsis Introduction Basic Algorithm Example Performance Applications

3 The Basics of Differential Evolution Stochastic, population-based optimisation algorithm Introduced by Storn and Price in 996 Developed to optimise real parameter, real valued functions General problem formulation is: For an objective function f : X R D R where the feasible region X, the minimisation problem is to find x X such that f(x ) f(x) x X where: f(x )

4 Why use Differential Evolution? Global optimisation is necessary in fields such as engineering, statistics and finance But many practical problems have objective functions that are nondifferentiable, non-continuous, non-linear, noisy, flat, multi-dimensional or have many local minima, constraints or stochasticity Such problems are difficult if not impossible to solve analytically DE can be used to find approximate solutions to such problems

5 Evolutionary Algorithms DE is an Evolutionary Algorithm This class also includes Genetic Algorithms, Evolutionary Strategies and Evolutionary Programming Initialisation Mutation Recombination Selection Figure : General Evolutionary Algorithm Procedure

6 Notation Suppose we want to optimise a function with D real parameters We must select the size of the population N (it must be at least ) The parameter vectors have the form: x i,g = [x,i,g, x,i,g,... x D,i,G ] i =,,..., N. where G is the generation number.

7 Initialisation Initialisation Mutation Recombination Selection Define upper and lower bounds for each parameter: x L j x j,i, x U j Randomly select the initial parameter values uniformly on the intervals [x L j, xu j ]

8 Mutation Initialisation Mutation Recombination Selection Each of the N parameter vectors undergoes mutation, recombination and selection Mutation expands the search space For a given parameter vector x i,g randomly select three vectors x r,g, x r,g and x r,g such that the indices i, r, r and r are distinct

9 Mutation Initialisation Mutation Recombination Selection Add the weighted difference of two of the vectors to the third v i,g+ = x r,g + F (x r,g x r,g ) The mutation factor F is a constant from [, ] v i,g+ is called the donor vector

10 Recombination Initialisation Mutation Recombination Selection Recombination incorporates successful solutions from the previous generation The trial vector u i,g+ is developed from the elements of the target vector, x i,g, and the elements of the donor vector, v i,g+ Elements of the donor vector enter the trial vector with probability CR

11 Recombination Initialisation Mutation Recombination Selection u j,i,g+ = v j,i,g+ x j,i,g if rand j,i CR or j = I rand if rand j,i > CR and j I rand i =,,..., N; j =,,..., D rand j,i U[, ], I rand is a random integer from [,,..., D] I rand ensures that v i,g+ x i,g

12 Selection Initialisation Mutation Recombination Selection The target vector x i,g is compared with the trial vector v i,g+ and the one with the lowest function value is admitted to the next generation u i,g+ if f(u i,g+ ) f(x i,g ) x i,g+ = i =,,..., N otherwise x i,g Mutation, recombination and selection continue until some stopping criterion is reached

13 Example: Ackley s function DE with N =, F =. and CR =. Ackley s function ( ) f(x, x ) = +e exp. n (x + x ) exp Find x [, ] such that f(x ) f(x) x [, ] f(x ) = ; x = (, ) ( ) n (cos(πx ) + cos(πx ))

14 Example: Ackley s function f(x) x x

15 Example: Ackley s function 8 x 6 x

16 Example: Initialisation x x

17 Example: Mutation x x

18 Example: Mutation x x

19 Example: Mutation x x

20 Example: Mutation x x

21 Example: Mutation x x

22 Example: Recombination x x

23 Example: Selection x x

24 Example: Selection x x

25 Example: Selection x x

26 Example: Mutation x x

27 Example: Mutation x x

28 Example: Mutation x x

29 Example: Mutation x x

30 Example: Mutation x x

31 Example: Recombination x x

32 Example: Recombination x x

33 Example: Selection x x

34 Example: Selection x x

35 Example: Generation x x

36 Example: Generation x x

37 Example: Movie Thirty generations of DE N =, F =. and CR =. Ackley s function

38 Performance There is no proof of convergence for DE However it has been shown to be effective on a large range of classic optimisation problems In a comparison by Storn and Price in 997 DE was more efficient than simulated annealing and genetic algorithms Ali and Törn () found that DE was both more accurate and more efficient than controlled random search and another genetic algorithm In Lampinen and Storn demonstrated that DE was more accurate than several other optimisation methods including four genetic algorithms, simulated annealing and evolutionary programming

39 Recent Applications Design of digital filters Optimisation of strategies for checkers Maximisation of profit in a model of a beef property Optimisation of fermentation of alcohol

40 Further Reading Price, K.V. (999), An Introduction to Differential Evolution in Corne, D., Dorigo, M. and Glover, F. (eds), New Ideas in Optimization, McGraw- Hill, London. Storn, R. and Price, K. (997), Differential Evolution - A Simple and Efficient Heuristic for Global Optimization over Continuous Spaces, Journal of Global Optimization,, pp. 9.