1. Fast Evolutionary Optimisation Temi avanzati di Intelligenza Artificiale - Lecture 6 Prof. Vincenzo Cutello Department of Mathematics and Computer Science University of Catania

2. Fast Evolutionary Optimisation - Lecture 52 Exercise Sheet  Function Optimisation by Evolutionary Programming  Fast Evolutionary Programming  Computational Studies  Summary

3. Fast Evolutionary Optimisation - Lecture 53 Global Optimisation

4. Fast Evolutionary Optimisation - Lecture 54 Global Optimisation by Mutation- Based EAs 1. Generate the initial population of  individuals, and set k = 1. Each individual is a real-valued vector, (x i ). 2. Evaluate the fitness of each individual. 3. Each individual creates a single offspring: for j = 1,…, n, where xi(j) denotes the j-th component of the vectors x i. N(0,1) denotes a normally distributed one-dimensional random number with mean zero and standard deviation one. N j (0,1) indicates that the random number is generated anew for each value of j. 4. Calculate the fitness of each offspring. 5. For each individual, q opponents are chosen randomly from all the parents and offspring with an equal probability. For each comparison, if the individual's fitness is no greater than the opponent's, it receives a “win”.

5. Fast Evolutionary Optimisation - Lecture 55 6. Select the  best individuals (from 2  ) that have the most wins to be the next generation. 7. Stop if the stopping criterion is satisfied; otherwise, k = k + 1 and go to Step 3.

6. Fast Evolutionary Optimisation - Lecture 56 Why N(0,1)?  The standard deviation of the Normal distribution determines the search step size of the mutation. It is a crucial parameter.  Unfortunately, the optimal search step size is problem-dependent.  Even for a single problem, different search stages require different search step sizes.  Self-adaptation can be used to get around this problem partially.

7. Fast Evolutionary Optimisation - Lecture 57 Function Optimisation by Classical EP (CEP) EP = Evolutionary Programming 1. Generate the initial population of  individuals, and set k = 1. Each individual is taken as a pair of real-valued vectors, (x i,  i ),  i  {1,…,  }. 2. Evaluate the fitness score for each individual (x i,  i ),  i  {1,…,  } of the population based on the objective function, f(x i ). 3. Each parent (x i,  i ), i = 1,…, , creates a single offspring (x’ i,  ’ i ), by: for j = 1,…,n where x j (j), x j ’ (j),  j (j) and  j ’ (j) denote the j-th component of the vectors x j, x j ’,  j and  j ’, respectively. N(0,1) denotes a normally distributed one-dimensional random number with mean zero and standard deviation one. N j (0,1) indicates that the random number is generated anew for each value of j. The factors  and  ’ have commonly set to

8. Fast Evolutionary Optimisation - Lecture 58 4. Calculate the fitness of each offspring (x i ’,  i ’ ),  i  {1,…,  }. 5. Conduct pairwise comparison over the union of parents (x i,  i ), and offspring (x i ’,  I ’ ),  i  {1,…,  }. For each individual, q opponents are chosen randomly from all the parents and offspring with an equal probability. For each comparison, if the individual's fitness is no greater than the opponent's, it receives a “win.” 6. Select the  individuals out of (x i,  i ), and (x i ’,  i ’ ),  i  {1,…,  } that have the most wins to be parents of the next generation. 7. Stop if the stopping criterion is satisfied; otherwise, k = k + 1 and go to Step 3.

9. Fast Evolutionary Optimisation - Lecture 59  What Do Mutation and Self-Adaptation Do

10. Fast Evolutionary Optimisation - Lecture 510 Fast EP  The idea comes from fast simulated annealing.  Use a Cauchy, instead of Gaussian, random number in Eq.(1) to generate a new offspring. That is, where  j is an Cauchy random number variable with the scale parameter t = 1, and is generated anew for each value of j.  Everything else, including Eq.(2), are kept unchanged in order to evaluate the impact of Cauchy random numbers.

11. Fast Evolutionary Optimisation - Lecture 511 Cauchy Distribution  Its density function is  where t > 0 is a scale parameter. The corresponding distribution function is

12. Fast Evolutionary Optimisation - Lecture 512 Gaussian and Cauchy Density Functions

13. Fast Evolutionary Optimisation - Lecture 513 Test Functions  23 functions were used in our computational studies. They have different characteristics.  Some have a relatively high dimension.  Some have many local optima.

14. Fast Evolutionary Optimisation - Lecture 514

15. Fast Evolutionary Optimisation - Lecture 515 Experimental Setup  Population size 100.  Competition size 10 for selection.  All experiments were run 50 times, i.e., 50 trials.  Initial populations were the same for CEP and FEP.

16. Fast Evolutionary Optimisation - Lecture 516 Experiments on Unimodal Functions  The value of t with 49 degrees of freedom is significant at  = 0,05 by a two-tailed test.

17. Fast Evolutionary Optimisation - Lecture 517 Discussions on Unimodal Functions  FEP performed better than CEP on f 3 -f 7.  CEP was better for f 1 and f 2.  FEP converged faster, even for f 1 and f 2 (for a long period).

18. Fast Evolutionary Optimisation - Lecture 518 Experiments on Multimodal Functions f 8 -f 13  The value of t with 49 degrees of freedom is significant at  = 0,05 by a two-tailed test.

19. Fast Evolutionary Optimisation - Lecture 519 Discussions on Multimodal Functions f 8 -f 13  FEP converged faster to a better solution.  FEP seemed to deal with many local minima well.

20. Fast Evolutionary Optimisation - Lecture 520 Experiments on Multimodal Functions f 14 -f 23  The value of t with 49 degrees of freedom is significant at  = 0,05 by a two-tailed test.

21. Fast Evolutionary Optimisation - Lecture 521 Discussions on Multimodal Functions f 14 -f 23  The results are mixed!  FEP and CEP performed equally well on f 16 and f 17. They are comparable on f 15 and f 18 – f 20.  CEP performed better on f 21 – f 23 (Shekel functions).  Is it because the dimension was low so that CEP appeared to be better?

22. Fast Evolutionary Optimisation - Lecture 522 Experiments on Low-Dimensional f 8 -f 13  The value of t with 49 degrees of freedom is significant at  = 0,05 by a two-tailed test.

23. Fast Evolutionary Optimisation - Lecture 523 Discussions on Low-Dimensional f 8 -f 13  FEP still converged faster to better solutions.  Dimensionality does not play a major role in causing the difference between FEP and CEP.  There must be something inherent in those functions which caused such difference.

24. Fast Evolutionary Optimisation - Lecture 524 The Impact of Parameter t on FEP | Part I  For simplicity, t = 1 in all the above experiments for FEP. However, this may not be the optimal choice for a given problem.  Table 1: The mean best solutions found by FEP using different scale parameter t in the Cauchy mutation for functions f 1 (1500), f 2 (2000), f 10 (1500), f 11 (2000), f 21 (100), f 22 (100) and f 23 (100). The values in “()” indicate the number of generations used in FEP. All results have been averaged over 50 runs.

25. Fast Evolutionary Optimisation - Lecture 525 The Impact of Parameter t on FEP | Part II  Table 2: The mean best solutions found by FEP using different scale parameter t in the Cauchy mutation for functions f 1 (1500), f 2 (2000), f 10 (1500), f 11 (2000), f 21 (100), f 22 (100) and f 23 (100). The values in “()” indicate the number of generations used in FEP. All results have been averaged over 50 runs.

26. Fast Evolutionary Optimisation - Lecture 526 Why Cauchy Mutation Performed Better  Given G(0,1) and C(1), the expected length of Gaussian and Cauchy jumps are:  It is obvious that Gaussian mutation is much localised than Cauchy mutation.

27. Fast Evolutionary Optimisation - Lecture 527 Why and When Large Jumps Are Beneficial  (Only one dimensional case is considered here for convenience's sake.)  Take the Gaussian mutation with G(0,  2 ) distribution as an example, i.e.,  the probability of generating a point in the neighbourhood of the global optimum x * is given by :  where  > 0 is the neighbourhood size and  is often regarded as the step size of the Gaussian mutation. Figure 4 illustrates the situation.

28. Fast Evolutionary Optimisation - Lecture 528  Figure 4: Evolutionary search as neighbourhood search, where x * is the global  optimum and  > 0 is the neighbourhood size.

29. Fast Evolutionary Optimisation - Lecture 529 An Analytical Result  It can be shown that when x* -  +  > . That is, the larger  is, the larger if x* -  +  > .  On the other hand, if x* -  +  > , then which indicates that decreases, exponentially, as  increases.

30. Fast Evolutionary Optimisation - Lecture 530 Empirical Evidence I  Table 3: Comparison of CEP's and FEP's final results on f 21 when the initial population is generated uniformly at random in the range of 0  x i  10 and 2.5  x i  5.5. The results were averaged over 50 runs. The number of generations for each run was 100.

31. Fast Evolutionary Optimisation - Lecture 531 Empirical Evidence II  Table 4: Comparison of CEP's and FEP's final results on f 21 when the initial population is generated uniformly at random in the range of 0  x i  10 and 0  x i  100 and a i 's were multiplied by 10. The results were averaged over 50 runs. The number of generations for each run was 100.

32. Fast Evolutionary Optimisation - Lecture 532 Summary  Cauchy mutation performs well when the global optimum is far away from the current search location. Its behaviour can be explained theoretically and empirically.  An optimal search step size can be derived if we know where the global optimum is. Unfortunately, such information is unavailable for real-world problems.  The performance of FEP can be improve by a set of more suitable parameters, instead of copying CEP's parameter setting.  Reference X. Yao, Y. Liu and G. Lin, “Evolutionary programming made faster,” IEEE Transactions on Evolutionary Computation, 3(2):82-102, July 1999.