1 / 59

PPT Primer on Evolutionary Algorithms (also known as Genetic Algorithms)

PPT Primer on Evolutionary Algorithms (also known as Genetic Algorithms). Introduction to Evolutionary Computation. Natural Evolution Search and Optimisation Hillclimbing / Gradient descent Local Search Population-Based Algorithms (i.e. Evolutionary Algorithms)

Download Presentation

PPT Primer on Evolutionary Algorithms (also known as Genetic Algorithms)

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. PPT Primer on Evolutionary Algorithms (also known as Genetic Algorithms)

  2. Introduction to Evolutionary Computation • Natural Evolution • Search and Optimisation • Hillclimbing / Gradient descent • Local Search • Population-Based Algorithms (i.e. Evolutionary Algorithms) • Advantages and Disadvantages of EAs • Applications of EAs • Reading Material and Resources

  3. Natural Evolution as a Problem Solving Method We seem to have evolved from tiny stuff in the sea. How??? The theory is: given: • a population of organisms which have a lifetime and which can reproduce in a challenging/changing environment • a way of continually generating diversity in new `child’ organisms A `survival of the fittest principle will naturally emerge: organisms which tend to have healthy, fertile children will dominate (I.e. their descendents will).

  4. Evolution/Survival of the Fittest In particular, any new mutation that appears in a child (e.g. longer neck, longer legs, thicker skin, longer gestation period, bigger brain, light-sensitive patch on the skin, a harmless `loose’ bone, etc etc) and which helps it in its efforts to survive long enough to have children, will become more and more widespread in future generations. The theory of evolution is the statement that all species on Earth have arisen in this way by evolution from one or more very simple self-reproducing molecules in the primeval soup. I.e. we have evolved via the accumulation of countless advantageous (in context) mutations over countless generations, and species have diversified to occupy niches, as a result of different environments favouring different mutations.

  5. Evolution as a Problem Solving Method Can view evolution as a way of solving the problem: How can I survive in this environment? The basic method of it is trial and error. I.e. evolution is in the family of methods that do something like this: 1. Come up with a new solution by randomly changing an old one. Does it work better than previous solutions? If yes, keep it and throw away the old ones. Otherwise, discard it. 2. Go to 1. But this appears to be a recipe for problem solving algorithms which take forever, with little or no eventual success!

  6. The Magic Ingredients Not so – since there are two vital things (and one other sometimes useful thing) we learn from natural evolution, which, with a sprinkling of our own commonsense added, lead to generally superb problem solving methods called evolutionary algorithms: Lesson1: Keep a population/collection of different things on the go. Lesson2: Select `parents’ with a relatively weakbias towards the fittest. It’s not really plain survival of the fittest, what works is the fitter you are, the more chance you have to reproduce, and it works best if even the least fit still have some chance. Lesson3: It can sometimes help to use recombination of two or more `parents’ – I.e. generate new candidate solutions by combining bits and pieces from different previous solutions.

  7. A Generic Evolutionary Algorithm Suppose you have to find a solution to some problem or other, and suppose, given any candidate solution s you have a function f(s) which measures how good s is as a solution to your problem. Generate an initial population P of randomly generated solutions (this is typically 100 or 500 or so). Evaluate the fitness of each. Then: Repeat until a termination condition is reached: • Selection: Choose some of P to be parents • Variation: Apply genetic operators to the parents to produce some children, and then evaluate the fitness of the children. • Population update: Update the population P by retaining some of the children and removing some of the incumbents.

  8. Basic Varieties of Evolutionary Algorithm • Selection: Choose some of P to be parents • Variation: Apply genetic operators … • Population update: Update the population P by … There are many different ways to select – e.g. choose top 10% of the population; choose with probability proportionate to fitness; choose randomly from top 20%, etc … There are many different ways to do this, and it depends much on the encoding (see next slide). We will learn certain standard ways. There are many several ways to do this, e.g. replace entire population with the new children; choose best |P| from P and the new ones, etc.

  9. Some of what EA-ists (theorists and practitioners) are Most concerned with: How to select? Always select the best? Bad results, quickly Select almost randomly? Great results, too slowly How to encode? Can make all the difference, and is intricately tied up with : How to vary? (mutation, recombination, etc…) small-step mutation preferred, recombination seems to be a principled way to do large steps, but large steps are usually abysmal. What parameters? How to adapt with time?

  10. What are they good for ? Suppose we want the best possible schedule for a university lecture timetable. • Or the best possible pipe network design for a ship’s engine room • Or the best possible design for an antenna with given requirements • Or a formula that fits a curve better than any others • Or the best design for a comms network in terms of reliability for given cost • Or the best strategy for flying a fighter aircraft • Or the best factory production schedule we can get, • Or the most accurate neural network for a data mining or control problem, • Or the best treatment plan (beam shapes and angles) for radiotherapy cancer treatment • And so on and so on ….! • The applications cover all of optimisation and machine learning.

  11. More like selective breeding than natural evolution Time

  12. Every Evolutionary Algorithm Given a problem to solve, a way to generate candidate solutions, and a way to assign fitness values: • Generate and evaluate a population of candidate solutions • Select a few of them • Breed the selected ones to obtain some new candidate solutions, and evaluate them • Throw out some of the population to make way for some of the new children. • Go back to step 2 until finished.

  13. Initial population

  14. Select

  15. Crossover

  16. Another Crossover

  17. A mutation

  18. Another Mutation

  19. Old population + children

  20. New Population: Generation 2

  21. Generation 3

  22. Generation 4, etc …

  23. Bentley.s thesis work Fixed wheel positions, constrained bounding area, Chromosome is a series of slices fitnesses evaluated via a simple airflow simulation

  24. Buy it

  25. The Evolutionary Computation Fossil Record The first published ideas using evolution in optimisation came in the 50s. But the lineage of current algorthms is like this: An intellectual curiosity Rechenberg, Berlin, Evolutionsstrategie Holland, Michigan Classifier Systems, Genetic plans Fogel, San Diego Evolutionary Programming 60s 80s 90s Goldberg, Michigan Genetic Algorithms A gift from Heaven Koza, Stanford Genetic Programming Ross, Corne, logistics Parmee, Eng. design Fleming, control systems Savic, Walters, Water systems

  26. One of the very first applications. Determine the internal shape of a two-phase jet nozzle that can achieve the maximum possible thrust under given starting conditions Ingo Rechenberg was the very first, with pipe-bend design This is slightly later work in the same lab, by Schwefel Starting point EA (ES) running Result A recurring theme: design freedom  entirely new and better designs based on principles we don’t yet understand.

  27. Some extra slides if time, illustrating some high-profile EAs An innovative EC-designed Propellor from Evolgics GmbH, Associated with Rechenberg’s group.

  28. Evolving Top Gun strategies

  29. Evolving Top Gun strategies

  30. Credit Jason Lohn NASA ST5 Mission had challenging requirements for antenna of 3 small spacecraft. EA designs outperformed human expert ones and are nearly spacebound.

  31. Credit Jason Lohn Oh no, we knew something like this would happen

  32. A Standard Evolutionary Algorithm The algorithm whose pseudocode is on the next slide is a steady state, replace-worst EA with tournament selection, using mutation, but no crossover. Parameters are popsize, tournament size, mutation-rate. It can be applied to any problem; the details glossed over are all problem-specific.

  33. A steady state, mutation-only, replace-worst EA with tournament selection 0. Initialise: generate a population of popsize random solutions, evaluate their fitnesses. • Run Select to obtain a parent solution X. • With probability mute_rate,mutate a copy of X to obtain a mutant M (otherwise M = X) • Evaluate the fitness of M. • Let W be the current worst in the population (BTR). If M is not less fit than W, then replace W with M. (otherwise do nothing) • If a termination condition is met (e.g. we have done 10,000 evals) then stop. Otherwise go to 2. Select: randomly choose tsize individuals from the population. Let c be the one with best fitness (BTR); return X.

  34. A generational, elitist,crossover+mutation EA with Rank-Based selection 0. Initialise: generate a population G of popsize random solutions, evaluate their fitnesses. • Run Select 2*(popsize – 1) times to obtain a collection I of 2*(popsize-1) parents. • Randomly pair up the parents in I (into popsize – 1 pairs) and apply Vary to produce a child from each pair. Let the set of children be C. • Evaluate the fitness of each child. • Keep the best in the population G (BTR) and delete the rest. • Add all the children to G. • If a termination condition is met (e.g. we have done 100 or more generations (runs through steps 1—5) then stop. Otherwise go to 1,

  35. A generational, elitist,crossover+mutation EA with Rank-Based selection, continued … Select: sort the contents of G from best to worst, assigning rank popsize to the best, popsize-1 to the next best, etc …, and rank 1 to the worst. The ranks sum to F = popsize(popsize+1)/2 Associate a probability Rank_i/F with each individual i. Using these probabilities, choose one individual X, and return X. Vary: 1. With probability cross_rate, do a crossover: I.e produce a child by applying a crossover operator to the two parents. Otherwise, let the child be a randomly chosen one of the parents. 2. Apply mutation to the child. 3. Return the mutated child.

  36. Back to Basics With your thirst for seeing example EAs temporarily quenched, the story now skips to simpler algorithms. This will help to explain what it is about the previous ones which make them work.

  37. The Travelling Salesperson Problem An example (hard) problem, for illustration The Travelling Salesperson Problem Find the shortest tour through the cities. The one below is length: 33 B C E D A

  38. Hillclimbing • 0. Initialise: Generate a random solution c; evaluate its • fitness, f(c). Call c the current solution. • 1. Mutate a copy of the current solution – call the mutant m • Evaluate fitness of m, f(m). • 2. If f(m) is no worsethan f(c), then replace c with m, • otherwise do nothing (effectively discarding m). • 3. If a termination condition has been reached, stop. • Otherwise, go to 1. Note. No population (well, population of 1). This is a very simple version of an EA, although it has been around for much longer.

  39. Why “Hillclimbing”? Suppose that solutions are lined up along the x axis, and that mutation always gives you a nearby solutions. Fitness is on the y axis; this is a landscape 9 6 10 7 5, 8 4 3 1 2 • Initial solution; 2. rejected mutant; 3. new current solution, • 4. New current solution; 5. new current solution; 6. new current soln • 7. Rejected mutant; 8. rejected mutant; 9. new current solution, • 10. Rejected mutant, …

  40. Here is our initial random solution ABDEC with fitness 32 Example: HC on the TSP We can encode a candidate solution to the TSP as a permutation Current solution Mutant B B C C E D E A D A

  41. HC on the TSP We randomly mutate (swap randomly chosen adjacent nodes) current to: ABEDC which has fitness 33 -- so current stays the same Current solution Mutant B B C C E D E A D A

  42. HC on the TSP We randomly mutate (swap randomly chosen adjacent nodes) current (ABDEC) to CBDEA which has fitness 38 -- so current stays the same Current solution Mutant B B C C E D E A D A

  43. HC on the TSP We randomly mutate (swap randomly chosen adjacent nodes) current (ABDEC) to BADEC which has fitness 28 So this becomes the new current solution Current solution Mutant B B C C E D E A D A

  44. HC on the TSP We randomly mutate (swap randomly chosen adjacent nodes) current (BADEC) to BADCE which also has fitness 28 This becomes the new current solution Current solution Mutant B B C C E D E A D A

  45. Landscapes f(s) Recall S, the search space, and f(s), the fitness of a candidate in S members of S lined up along here The structure we get by imposing f(s) on S is called a landscape What does the landscape look like if f(s) is a random number generator? What kinds of problems would have very smooth landscapes? What is the importance of the mutation operator in all this?

  46. Neighbourhoods Let s be an individual in S, f(s) is our fitness function, and M is our mutation operator, so that M(s1)  s2, where s2 is a mutant of s1. Given M, we can usually work out the neighbourhood of an individual point s – the neighbourhood of s is the set of all possible mutants of s E.g. Encoding: permutations of k objects (e.g. for k-city TSP) Mutation: swap any adjacent pair of objects. Neighbourhood: Each individual has k neighbours. E.g. neighbours of EABDC are: {AEBDC, EBADC, EADBC, EABCD, CABDE} Encoding: binary strings of length L (e.g. for L-item bin-packing) Mutation: choose a bit randomly and flip it. Neighbourhood: Each individual has L neighbours. E.g. neighbours of 00110 are: {10110, 01110, 00010, 00100, 00111}

  47. Landscape Topology Mutation operators lead to slight changes in the solution, which tend to lead to slight changes in fitness. I.e. the fitnesses in the neighbourhood of s are often similar to the fitness of s. Landscapes tend to be locally smooth What about big mutations ?? It turns out that ….

  48. Typical Landscapes f(s) members of S lined up along here Typically, with large (realistic) problems, the huge majority of the landscape has very poor fitness – there are tiny areas where the decent solutions lurk. So, big random changes are very likely to take us outside the nice areas.

  49. Typical Landscapes II Plateau Unimodal Multimodal Deceptive As we home in on the good areas, we can identify broad types of Landscape feature. Most landscapes of interest are predominantly multimodal. Despite being locally smooth, they are globally rugged

  50. Beyond Hillclimbing HC clearly has problems with typical landscapes: There are two broad ways to improve HC, from the algorithm viewpoint: • Allow downhill moves – a family of methods called Local Search does this in various ways. • Have a population – so that different regions can be explored inherently in parallel – I.e. we keep `poor’ solutions around and give them a chance to `develop’.

More Related