1. [Start] Generate random population of N chromosomes (suitable solutions for the problem) 2. [Fitness] Evaluate the fitness f(x) of each chromosome x in the population 3. [New population] Create a new population by repeating following steps until the new population is complete 1. [Selection] Select two parent chromosomes from a population according to their fitness (the better fitness, the bigger chance to be selected) 2. [Crossover] With a crossover probability cross over the parents to form a new offspring (children). If no crossover was performed, offspring is an exact copy of parents. 3. [Mutation] With a mutation probability mutate new offspring at each locus (position in chromosome). 4. [Accepting] Place new offspring in a new population
46
Embed
[Start] Generate random population of N chromosomes (suitable solutions for the problem)
[Start] Generate random population of N chromosomes (suitable solutions for the problem) [Fitness] Evaluate the fitness f(x) of each chromosome x in the population [New population] Create a new population by repeating following steps until the new population is complete - PowerPoint PPT Presentation
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
1. [Start] Generate random population of N chromosomes (suitable solutions for the problem)
2. [Fitness] Evaluate the fitness f(x) of each chromosome x in the population 3. [New population] Create a new population by repeating following steps
until the new population is complete 1. [Selection] Select two parent chromosomes from a population
according to their fitness (the better fitness, the bigger chance to be selected)
2. [Crossover] With a crossover probability cross over the parents to form a new offspring (children). If no crossover was performed, offspring is an exact copy of parents.
3. [Mutation] With a mutation probability mutate new offspring at each locus (position in chromosome).
4. [Accepting] Place new offspring in a new population 4. [Replace] Use new generated population for a further run of algorithm 5. [Test] If the end condition is satisfied, stop, and return the best solution in
current population 6. [Loop] Go to step 2.
• Check the following applets– http://www.obitko.com/tutorials/genetic-algorithms/exam
• The GA described so far is similar to Holland’s original GA.
• It is now known as the simple genetic algorithm (SGA).
• Other GAs use different:– Representations– Mutations– Crossovers– Selection mechanisms
Crossover enhancements
• Multipoint crossover
• Uniform crossover
Crossover OR mutation?
• Decade long debate: which one is better• Answer (at least, rather wide
agreement):– it depends on the problem, but– in general, it is good to have both– these two have different roles– a mutation-only-GA is possible, an xover-
only-GA would not work
• Crossover is explorative– Discovering promising areas in the search space.– It makes a big jump to an area somewhere “in
between” two (parent) areas.• Mutation is exploitative – Optimizing present information within an already
discovered promising region.– Creating small random deviations and thereby not
wandering far from the parents.
• They complement each other.– Only crossover can bring together information
from both parents.– Only mutation can introduce completely new
information.
Feature Selection, Feature Extraction
Why to do it?1. We’re interested in features – we want to
know which are relevant. If we fit a model, it should be interpretable.• facilitate data visualization and data understanding• reduce experimental costs (measurements)
2. We’re interested in prediction – features are not interesting in themselves, we just want to build a good predictor.• faster training• defy the curse of dimensionality
Selection vs. Extraction• In feature selection we try to find the best subset of the
input feature set.
• In feature extraction we create new features based on transformation or combination of the original feature set.
• Both selection and extraction lead to the dimensionality reduction.
• No clear cut evidence that one of them is superior to the other on all types of task.
Feature selection(FS)
Classification of FS methods• Filter– Assess the relevance of features only by looking at the intrinsic
properties of the data.– Usually, calculate the feature relevance score and remove low-
scoring features.• Wrapper– Bundle the search for best model with the FS.– Generate and evaluate various subsets of features. The
evaluation is obtained by training and testing a specific ML model.
• Embedded– The search for an optimal subset is built into the classifier
construction (e.g. decision trees).
Filter methods
• Two steps (score-and-filter approach)1. assess each feature individually for ist potential in
discriminating among classes in the data2. features falling beyong threshold are eliminated
• Advantages:– easily scale to high-dimensional data– simple and fast– independent of the classification algorithm
• Disadvantages:– ignore the interaction with the classifier– most techniques are univariate (each feature is considered
separately)
Scores in filter methods• Distance measures– Euclidean distance
• Information measures– information gain– mutual information
• complexity: O(d)
Wrappers
• Search for the best feature subset in combination with a fixed classification method.
• The goodness of a feature subset is determined using cross-validation (k-fold, LOOCV)
• Advantages:– interaction between feature subset and model selection– take into account feature dependencies– generally more accurate
• Disadvantages:– higher risk of overfitting than filter methods– very computationally intensive
Exhaustive search• Evaluate all possible subsets using exhaustive search
– this leads to the optimum subset.• For a total of d variables, and subset of size p, the
total number of possible subsets is
• complexity: O(2d) (exponential)• Various strategies how to reduce the search space.– They are still O(2d), but much faster (at least 1000-times)– e.g. “branch and bound”
!! !
dd p p
e.g. d = 100, p = 10 → ≈2×1013
Stochastic
• Genetic algorithms• Simulated Annealing
Deterministic
• Sequential Forward Selection (SFS)• Sequential Backward Selection (SBS)• “ Plus q take away r ” Selection• Sequential Forward Floating Search (SFFS)• Sequential Backward Floating Search (SBFS)
Sequential Forward Selection
• SFS• At the beginning select the best feature using
a scalar criterion function.• Add one feature at a time which along with
already selected features maximizes the criterion function.
• A greedy algorithm, cannot retract (also called nesting effect).
• Complexity is O(d)
Sequential Backward Selection
• SBS• At the beginning select all d features.• Delete one feature at a time and select the
subset which maximize the criterion function.• Also a greedy algorithm, cannot retract.• Complexity is O(d).
“Plus q take away r” Selection
• At first add q features by forward selection, then discard r features by backward selection
• Need to decide optimal q and r• No subset nesting problems Like SFS and SBS
Sequential Forward Floating Search
• SFFS• It is a generalized “plus q take away r”
algorithm• The value of q and r are determined
automatically• Close to optimal solution• Affordable computational cost • Also in backward disguise
Embedded FS
• The feature selection process is done inside the ML algorithm.
• Decision trees– In final tree, only a subset of features are used
• Regularization– It effectively “shuts down” unnecessary features.– Pruning in NN.
Feature extraction(FE)
• FS – indetify and select the “best” features with respect to the target task.– Selected features retain their original physical
interpretation.• FE – create new features as a transformation
(combination) of original features. Usually followed by FS.– May provide better discriminatory ability than the
best subset.– Do not retain the original physical interpretation,
may not have clear meaning.
Principal Component Analysis(PCA)
x1
x2
x1
x2
Make datato have zero mean (i.e. move data into [0, 0] point).
centering
x1
x2
This is a line given by equation w0 + w1x1 + w2x2
This is another linew’0 + w’1x1 + w’2x2
x1
x2
The variability in data is highest along this line. It is called 1st principal component.
And this is 2nd principal component.
x1
x2
Principal components (PC’s) are linear combinations of original coordinates.
The coefficients of linear combination (w0, w1, …) are called loadings.
In the transformed coordinate system, individual data points have different coordinates, these are called scores.
w0 + w1x1 + w2x2
w’0 + w’1x1 + w’2x2
• PCA - orthogonal linear transformation that changes the data into a new coordinate system such that the variance is put in order from the greatest to the least.
• Solve the problem = find new orthogonal coordinate system = find loadings
• PC’s (vectors) and their corresponding variances (scalars) are found by eigenvalue decompositions of the covariance matrix C = XXT of the xi variables.– Eigenvector corresponding to the largest eigenvalue is 1st PC.– The 2nd eigenvector (the 2nd largest eigenvalue) is orthogonal
to the 1st one. …• Eigenvalue decomposition is computed using standard
algorithms: eigen decomposition of covariance matrix (e.g. QR algorithm), SVD of mean centered data matrix.
Interpretation of PCA
• New variables (PCs) have a variance equal to their corresponding eigenvalue
Var(Yi)= i for all I = 1…p• Small i small variance data changes little
in the direction of component Yi
• The relative variance explained by each PC is given by i / i
How many components?
• Enough PCs to have a cumulative variance explained by the PCs that is >50-70%
• Kaiser criterion: keep PCs with eigenvalues >1• Scree plot: represents the ability of PCs to
explain de variation in data
Difficulty in Searching Global Optima
startingpoint
descenddirection
local minima
global minima
barrier to local search
Introduction to Simulated Annealing, Dr. Gildardo Sánchez ITESM Campus Guadalajara
Consequences of the Occasional Ascents
Help escaping the local optima.
desired effect
Might pass global optima after reaching it
adverse effect
Introduction to Simulated Annealing, Dr. Gildardo Sánchez ITESM Campus Guadalajara
Simulated annealing
• Slowly cool down a heated solid, so that all particles arrange in the ground energy state.
• At each temperature wait until the solid reaches its thermal equilibrium.
– Perturb (randomly) the current state to a new state– E is the difference in energy between current and new state– If E < 0 (new state is lower), accept new state as current state– If E 0 , accept new state with probability
• Eventually the systems evolves into thermal equilibrium at temperature T
• When equilibrium is reached, temperature T can be lowered and the process can be repeated
expB
EPk T
Metropolis, 1953Metropolis algorithm
Simulated annealing
• Same algorithm can be used for combinatorial optimization problems:– Energy E corresponds to the objective function C– Temperature T is parameter controlled within the
IF accept THEN update(config.j); UNTIL equilibrium is approached sufficient closely; T := next_lower(T);UNTIL system is frozen or stop criterion is reached
Parameters
• Choose the start value of T so that in the beginning nearly all perturbations are accepted (exploration), but not too big to avoid long run times
• At each temperature, search is allowed to proceed for a certain number of steps, L(k).
• The function next_lower (T(k)) is generally a simple function to decrease T, e.g. a fixed part (80%) of current T.
• At the end T is so small that only a very small number of the perturbations is accepted (exploitation).
• The choice of parameters {T(k), L(k)} is called the cooling schedule.
• If possible, always try to remember explicitly the best solution found so far; the algorithm itself can leave its best solution and not find it again.