Survey
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
Evolutionary Computational Inteliigence Lecture 6b: Towards Parameter Control Ferrante Neri University of Jyväskylä 1 Motivation 1 An EA has many strategy parameters, e.g. mutation operator and mutation rate crossover operator and crossover rate selection mechanism and selective pressure (e.g. tournament size) population size Good parameter values facilitate good performance Q1 How to find good parameter values ? 2 Motivation 2 EA parameters are rigid (constant during a run) BUT an EA is a dynamic, adaptive process THUS optimal parameter values may vary during a run Q2: How to vary parameter values? 3 Parameter tuning Parameter tuning: the traditional way of testing and comparing different values before the “real” run Problems: users mistakes in settings can be sources of errors or sub-optimal performance costs much time parameters interact: exhaustive search is not practicable good values may become bad during the run (e.g. Population size) 4 Parameter Setting: Problems 5 A wrong parameter setting can lead to an undesirable algorithmic behavious since it can lead to stagnation or premature convergence Too large population size, stagnation Too small population size, premature convergence In some “moments” of the evolution I would like to have a large pop size (when I need to explore and prevent premature convergence); in other “moments” I would like to have a small one (when I need to exploit available genotypes) Parameter control Parameter control: setting values on-line, during the actual run, I would like that the algorithm “decides” by itself how to properly vary parameter setting over the run Some popular options for pursuing this aim are: 6 predetermined time-varying schedule p = p(t) using feedback from the search process encoding parameters in chromosomes and rely on natural selection (similar to ES self-adaptation) Related Problems Problems: finding optimal p is hard, finding optimal p(t) is harder still user-defined feedback mechanism, how to ”optimize”? when would natural selection work for strategy parameters? Provisional answer: 7 In agreement with the No Free Lunch Theorem, optimal control strategy does not exist. Nevertheless, there are a plenty of interesting proposals that can be very performing in some problems. Some of these strategies are very problem oriented while some others are much more robust and thus applicable in a fairly wide spectrum of optimization problems How Three major types of parameter control: 8 deterministic: some rule modifies strategy parameter without feedback from the search (based on some counter) adaptive: feedback rule based on some measure monitoring search progress self-adaptative: parameter values evolve along with solutions; encoded onto chromosomes they undergo variation and selection Global taxonomy PARAMETER SETTING 9 PARAMETER TUNING PARAMETER CONTROL (before the run) (during the run) DETERMINISTIC ADAPTIVE SELF-ADAPTIVE (time dependent) (feedback from search) (coded in chromosomes) Evidence informing the change The parameter changes may be based on: time or nr. of evaluations (deterministic control) population statistics (adaptive control) – – – 10 progress made population diversity gene distribution, etc. relative fitness of individuals creeated with given values (adaptive or self-adaptive control) Evidence informing the change 11 Absolute evidence: predefined event triggers change, e.g. increase pm by 10% if population diversity falls under threshold x Direction and magnitude of change is fixed Relative evidence: compare values through solutions created with them, e.g. increase pm if top quality offspring came by high mut. Rates Direction and magnitude of change is not fixed Deterministic Control 12 It is based on a priori designed scheme which takes into account the variable time The general idea is that the algorithm at the beginning of the process has different needs compared to the end of it The main disadvantage of such an approach is that the algorithmic designer is supposed to know beforehand when the changes in parameter setting must be carried out Deterministic Control 13 1st example: the population is enlarged at the end of the optimization process (after a prearranged number of generations) 2nd example (Arabas 1994): assign to each solution a lifetime parameter. Fitter individuals must survive longer. The population is thus variable over a certain number of generations Adaptive Control 14 It is online measured an index to perform the parameter setting of the subsequent generations Fitness improvements: if the algorithmic response in terms of improvements are relevant, then a parameter is unchanged. When the parameter setting is inefficient then the parameter setting is modified (e.g. mutation rate increased when no improvements are found) Adaptive Control 15 Fitness diversity: in high diversity conditions the algorithm needs to exploit available genotypes, in low diversity conditions the algorithm needs to detect new genotypes and search directions A measurement of the diversity can be employed for enlarging population size in lo diversity condition and shrinking in high diversity conditions (analogously for mutation rate) Adaptive Control 1st example: min 1, f best f avg fbest f v S pop S pop S pop 1 pm pmmax 1 16 Adaptive Control 2nd example 1 f avg fbest f worst fbest f v S pop S pop S pop 1 pm pmmax 1 17 Adaptive Control 3rd example: f min 1, f avg f v S pop S pop S pop 1 pm pmmax 1 18 Self-adaptive control It encodes the control parameter within the genotype of the solution and it is based on the idea that it should evolve with the population x1 , x2 ,...xn , ctr _ par 19 Self-adaptive control 20 If for example the control parameter measures the improvements due to the mutation, it can affect the mutation rate Another popular option is to employ such a parameter for deciding the replication in the mating pool of the individuals (if the solution looks very promising) in order to have a tailored selection pressure Evaluation / Summary Parameter control offers the possibility to use appropriate values in various stages of the search Adaptive and self-adaptive parameter control – – – Robustness, insensivity of EA for variations assumed – – 21 offer users “liberation” from parameter tuning delegate parameter setting task to the evolutionary process the latter implies a double task for an EA: problem solving + self-calibrating (overhead) If no. of parameters is increased by using (self)adaptation For the “meta-parameters” introduced in methods