10 16 19
play

10/16/19 Parameter Control Genetic Algorithms Motivation - PDF document

Jan 03, 2023 •582 likes •659 views

10/16/19 Parameter Control Genetic Algorithms Motivation Parameter setting Tuning Control Examples Where to apply parameter control How to apply parameter control Parameter Control Motivation Motivation An EA

  1. 10/16/19 Parameter Control Genetic Algorithms • Motivation • Parameter setting – Tuning – Control • Examples • Where to apply parameter control • How to apply parameter control Parameter Control Motivation Motivation An EA has many parameters that affect the search, e.g. EA parameters are rigid (constant during a run) — mutation operator and mutation rate BUT — crossover operator and crossover rate an EA is a dynamic, adaptive process — selection mechanism and selective pressure ( e.g. THUS tournament size) optimal parameter values may vary during a run — population size Q2: How to vary parameter values? Good parameter values facilitate good performance Q1 How to find good parameter values ? 1

  2. 10/16/19 Parameter Settings: Parameter Setting 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 — parameters interact: exhaustive search is not practicable — costs much time even with “smart” tuning — good values may become bad during the run Parameter Settings: Examples: Control Varying mutation step size Parameter control: setting values on-line, during the Problem to solve: actual run, e.g. — min f(x 1 ,…,x n ) — L i £ x i £ U i — predetermined time-varying schedule p = p(t) for i = 1,…,n bounds — finding optimal p is hard, finding optimal p(t) is harder — g j (x) £ 0 for j = 1,…,q inequality constraints — h k (x) = 0 for k = q+1,…,m equality constraints — using feedback from the search process — still user-defined feedback mechanism, how to ``optimize"? Algorithm: — EA with real-valued representation x = (x 1 ,…,x n ) — encoding parameters in chromosomes and rely on — arithmetic averaging crossover selection — Gaussian mutation: x’ i = x i + N(0, s ) — will natural selection work for strategy parameters? standard deviation s is called mutation step size — how to implement effectively? 2

  3. 10/16/19 Examples: Examples: Varying mutation step size, option 1 Varying mutation step size, option 2 Replace the constant s by a function s (t) Replace the constant s by a function s (t) updated after every n steps by the 1/5 success rule: σ ( t )= 1 - 0 .9 × t T 0 £ t £ T is the current generation number 1/5 success rule (Rechenberg 1973): 1/5 of mutations should be successful – mutant more fit — Characteristics: than parent — changes in s are independent from the search progress — strong user control of s by the above formula — s is fully predictable — a given s acts on all individuals of the population Examples: Examples: Varying mutation step size, option 2 Varying mutation step size, option 3 Replace the constant s by a function s (t) updated after — Assign a personal s to each individual every n steps by the 1/5 success rule: — Incorporate this s into the chromosome: (x 1 , …, x n , s ) — Apply variation operators to x i ‘s and s ì s (t - n) / c if p > 0 . 2 σ ' = σ × e N( 0, σ ) s ï s s - × < (t) = (t n) c if p 0 . 2 0 < # < 1 í s x i ' = x i +N( 0, σ ' ) ï s - (t n) otherwise î — Characteristics: — Characteristics: — changes in s are results of natural selection — changes in s are based on feedback from the search progress — (almost) no user control of s — some user control of s by the above formula — s is not predictable — s is not predictable — a given s acts on one individual — a given s acts on all individuals of the population 3

  4. 10/16/19 Examples: Examples: Varying mutation step size, option 4 Varying penalties Constraints Assign a personal s to each variable in each individual Incorporate s ’s into the chromosomes: (x 1 , …, x n , s 1 , …, s n ) — g j (x) £ 0 for j = 1,…,q inequality constraints — h k (x) = 0 for k = q+1,…,m equality constraints Apply variation operators to x i ‘s and s i ‘s are handled by penalties: σ i ' = σ i × e N( 0, τ ) x i ' = x i +N( 0, σ i ' ) eval(x) = f(x) + W × penalty(x) — Characteristics: — changes in s i are results of natural selection where — (almost) no user control of s i ì 1 for violated constrai nt m — s i is not predictable å = penalty ( x ) í — a given s i acts on one gene of one individual 0 for satisfied constrai nt = î j 1 Examples: Examples: Varying penalties, option 1 Varying penalties, option 2 Replace the constant W by a function W(t) Replace the constant W by W(t) updated in each generation b ´ ì W(t) if last k champions all feasible ´ α W(t) = ( C t) ï g ´ W(t + 1 ) = í W(t) if last k champions all infeasible ï W(t) otherwise 0 £ t £ T is the current generation number î b < 1, g > 1, b ´ g ¹ 1 champion: best of its generation — Characteristics: — changes in W independent from the search progress — Characteristics: — strong user control of W by the above formula — changes in W are based on feedback from the search progress — some user control of W by the above formula — W is fully predictable — W is not predictable — a given W acts on all individuals of the population — a given W acts on all individuals of the population 4

  5. 10/16/19 Examples: Examples: Varying penalties, option 3 Lessons learned Assign a personal W to each individual in population Various forms of parameter control can be distinguished by: Incorporate this W into the chromosome: (x 1 , …, x n , W) Apply variation operators to W and each x i — primary features: — what component of the EA is changed Alert: — how the change is made eval ((x, W)) = f (x) + W × penalty(x) — secondary features: while for mutation step sizes we had — evidence/data backing up changes eval ((x, s )) = f (x) — level/scope of change this option is thus “cheating” Þ algorithm can improve the evaluation by evolving smaller weights W rather than improving f(x) Examples: Where to apply parameter control Lessons learned Various forms of parameter control can be distinguished by: Practically any EA component can be parameterized and thus controlled on-the-fly: — representation (x 1 , ..., x n , σ) (x 1 , …, x n , σ 1 , (x 1 , ..., x n , W) σ(t) = 1-0.9*t/T σ' = σ/c, if r > W(t) = (C*t) α W'= β *W, if b i ∈ F ⅕ ... …, σ n ) — evaluation function What Step size Step size Step Step Penalty Penalty Penalty — variation operators size size weight weight weight — selection operator (parent or mating selection) How Deterministic Adaptive Self- Self- Deterministic Adaptive Self- adaptive adaptive adaptive — replacement operator (survival or environmental Time (Fitness) (Fitness) Time Constraint (Fitness) Evidence Successful selection) mutations satisfaction rate history — population (size, topology) Scope Population Population Individual Gene Population Population Individual 5

  6. 10/16/19 How to apply parameter control How to apply parameter control Global taxonomy Three major types of parameter control: — 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 Evidence: Evidence: Informing the change Informing the change The parameter changes may be based on: — Absolute evidence: predefined event triggers change, e.g. increase p m by 10% if population diversity falls under — time or nr. of evaluations (deterministic control) threshold x — population statistics (adaptive control) — Direction and magnitude of change is fixed — progress made — population diversity — Relative evidence: compare values through solutions — gene distribution, etc. created with them, e.g. increase p m if top quality offspring came by high mutation rates — relative fitness of individuals created with given values — Direction and magnitude of change is not fixed (adaptive or self-adaptive control) 6

  7. 10/16/19 Evidence: Scope/level Refined taxonomy The parameter may take effect on different levels: l Combinations of types and evidences l Possible: + — environment (fitness function) l Impossible: - — population — individual — sub-individual Note: given component (parameter) determines possibilities Thus: scope/level is a derived or secondary feature in the classification scheme Evaluation/Summary — Parameter control offers the possibility to use appropriate values in various stages of the search — Adaptive and self-adaptive parameter control — 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) 7

Recommend

More recommend