Introduction to Statistical Machine Learning Marcus Hutter - - PowerPoint PPT Presentation

Introduction to Statistical Machine Learning

• 1 -

Marcus Hutter

Introduction to Statistical Machine Learning

Marcus Hutter

Canberra, ACT, 0200, Australia http://www.hutter1.net/ ANU RSISE NICTA

Machine Learning Summer School MLSS-2009, 26 Janurary – 6 February, Canberra

Introduction to Statistical Machine Learning

• 2 -

Marcus Hutter

Abstract

This course provides a brief overview of the methods and practice of statistical machine learning. It’s purpose is to (a) give a mini-introduction and background to logicians interested in the AI courses, and (b) to summarize the core concepts covered by the machine learning courses during this week. Topics covered include Bayesian inference and maximum likelihood modeling; regression, classification, density estimation, clustering, principal component analysis; parametric, semi-parametric, and non-parametric models; basis functions, neural networks, kernel methods, and graphical models; deterministic and stochastic optimization; overfitting, regularization, and validation.

Introduction to Statistical Machine Learning

• 3 -

Marcus Hutter

Table of Contents

• 1. Introduction / Overview / Preliminaries
• 2. Linear Methods for Regression
• 3. Nonlinear Methods for Regression
• 4. Model Assessment & Selection
• 5. Large Problems
• 6. Unsupervised Learning
• 7. Sequential & (Re)Active Settings
• 8. Summary

Intro/Overview/Preliminaries

• 4 -

Marcus Hutter

1 INTRO/OVERVIEW/PRELIMINARIES

• What is Machine Learning? Why Learn?
• Related Fields
• Applications of Machine Learning
• Supervised↔Unsupervised↔Reinforcement Learning
• Dichotomies in Machine Learning
• Mini-Introduction to Probabilities

Intro/Overview/Preliminaries

• 5 -

Marcus Hutter

What is Machine Learning?

Machine Learning is concerned with the development of algorithms and techniques that allow computers to learn Learning in this context is the process of gaining understanding by constructing models of observed data with the intention to use them for prediction.

Related fields

• Artificial Intelligence: smart algorithms
• Statistics: inference from a sample
• Data Mining: searching through large volumes of data
• Computer Science: efficient algorithms and complex models

Intro/Overview/Preliminaries

• 6 -

Marcus Hutter

Why ‘Learn’?

There is no need to “learn” to calculate payroll Learning is used when:

• Human expertise does not exist (navigating on Mars),
• Humans are unable to explain their expertise (speech recognition)
• Solution changes in time (routing on a computer network)
• Solution needs to be adapted to particular cases (user biometrics)

Example: It is easier to write a program that learns to play checkers or backgammon well by self-play rather than converting the expertise of a master player to a program.

Intro/Overview/Preliminaries

• 7 -

Marcus Hutter

Handwritten Character Recognition

an example of a difficult machine learning problem Task: Learn general mapping from pixel images to digits from examples

Intro/Overview/Preliminaries

• 8 -

Marcus Hutter

Applications of Machine Learning

machine learning has a wide spectrum of applications including:

• natural language processing,
• search engines,
• medical diagnosis,
• detecting credit card fraud,
• stock market analysis,
• bio-informatics, e.g. classifying DNA sequences,
• speech and handwriting recognition,
• object recognition in computer vision,
• playing games – learning by self-play: Checkers, Backgammon.
• robot locomotion.

Intro/Overview/Preliminaries

• 9 -

Marcus Hutter

Some Fundamental Types of Learning

• Supervised Learning

Classification Regression

• Unsupervised Learning

Association Clustering Density Estimation

• Reinforcement Learning

Agents

• Others

SemiSupervised Learning Active Learning

Intro/Overview/Preliminaries

• 10 -

Marcus Hutter

Supervised Learning

• Prediction of future cases:

Use the rule to predict the output for future inputs

• Knowledge extraction:

The rule is easy to understand

• Compression:

The rule is simpler than the data it explains

• Outlier detection:

Exceptions that are not covered by the rule, e.g., fraud

Intro/Overview/Preliminaries

• 11 -

Marcus Hutter

Classification

Example: Credit scoring Differentiating between low-risk and high-risk customers from their Income and Savings Discriminant: IF income > θ1 AND savings > θ2 THEN low-risk ELSE high-risk

Intro/Overview/Preliminaries

• 12 -

Marcus Hutter

Regression

Example: Price y = f(x)+noise of a used car as function of age x

Intro/Overview/Preliminaries

• 13 -

Marcus Hutter

Unsupervised Learning

• Learning “what normally happens”
• No output
• Clustering: Grouping similar instances
• Example applications:

Customer segmentation in CRM Image compression: Color quantization Bioinformatics: Learning motifs

Intro/Overview/Preliminaries

• 14 -

Marcus Hutter

Reinforcement Learning

• Learning a policy: A sequence of outputs
• No supervised output but delayed reward
• Credit assignment problem
• Game playing
• Robot in a maze
• Multiple agents, partial observability, ...

Intro/Overview/Preliminaries

• 15 -

Marcus Hutter

Dichotomies in Machine Learning

(machine) learning / statistical ⇔ logic/knowledge-based (GOFAI) induction ⇔ prediction ⇔ decision ⇔ action regression ⇔ classification independent identically distributed ⇔ sequential / non-iid

• nline learning

⇔

• ffline/batch learning

passive prediction ⇔ active learning parametric ⇔ non-parametric conceptual/mathematical ⇔ computational issues exact/principled ⇔ heuristic supervised learning ⇔ unsupervised ⇔ RL learning

Intro/Overview/Preliminaries

• 16 -

Marcus Hutter

Probability Basics

Probability is used to describe uncertain events; the chance or belief that something is or will be true. Example: Fair Six-Sided Die:

• Sample space: Ω = {1, 2, 3, 4, 5, 6}
• Events: Even= {2, 4, 6}, Odd= {1, 3, 5}

⊆ Ω

• Probability: P(6) = 1

6, P(Even) = P(Odd) = 1 2

• Outcome: 6 ∈ E.
• Conditional probability: P(6|Even) = P(6and Even)

P(Even) = 1/6 1/2 = 1 3 General Axioms:

• P({}) = 0 ≤ P(A) ≤ 1 = P(Ω),
• P(A ∪ B) + P(A ∩ B) = P(A) + P(B),
• P(A ∩ B) = P(A|B)P(B).

Intro/Overview/Preliminaries

• 17 -

Marcus Hutter

Probability Jargon

Example: (Un)fair coin: Ω = {Tail,Head} ≃ {0, 1}. P(1) = θ ∈ [0, 1]: Likelihood: P(1101|θ) = θ × θ × (1 − θ) × θ Maximum Likelihood (ML) estimate: ˆ θ = arg maxθ P(1101|θ) = 3

4

Prior: If we are indifferent, then P(θ) =const. Evidence: P(1101) =

θ P(1101|θ)P(θ) = 1 20 (actually

• )

Posterior: P(θ|1101) = P(1101|θ)P(θ)

P(1101)

∝ θ3(1 − θ) (BAYES RULE!). Maximum a Posterior (MAP) estimate: ˆ θ = arg maxθ P(θ|1101) = 3

4

Predictive distribution: P(1|1101) = P(11011)

P(1101) = 2 3

Expectation: E[f|...] =

θ f(θ)P(θ|...), e.g. E[θ|1101] = 2 3

Variance: Var(θ) = E[(θ − Eθ)2|1101] =

2 63

Probability density: P(θ) = 1

εP([θ, θ + ε]) for ε → 0

Linear Methods for Regression

• 18 -

Marcus Hutter

2 LINEAR METHODS FOR REGRESSION

• Linear Regression
• Coefficient Subset Selection
• Coefficient Shrinkage
• Linear Methods for Classifiction
• Linear Basis Function Regression (LBFR)
• Piecewise linear, Splines, Wavelets
• Local Smoothing & Kernel Regression
• Regularization & 1D Smoothing Splines

Linear Methods for Regression

• 19 -

Marcus Hutter

Linear Regression

fitting a linear function to the data

• Input “feature” vector x := (1 ≡ x(0), x(1), ..., x(d)) ∈ I

Rd+1

• Real-valued noisy response y ∈ I

R.

• Linear regression model:

ˆ y = fw(x) = w0x(0) + ... + wdx(d)

• Data: D = (x1, y1), ..., (xn, yn)
• Error or loss function:

Example: Residual sum of squares: Loss(w) = n

i=1(yi − fw(xi))2

• Least squares (LSQ) regression:

ˆ w = arg minw Loss(w)

• Example: Person’s weight y as a function of age x1, height x2.

Linear Methods for Regression

• 20 -

Marcus Hutter

Coefficient Subset Selection

Problems with least squares regression if d is large:

• Overfitting: The plane fits the data well (perfect for d ≥ n),

but predicts (generalizes) badly.

• Interpretation: We want to identify a small subset of

features important/relevant for predicting y. Solution 1: Subset selection: Take those k out of d features that minimize the LSQ error.

Linear Methods for Regression

• 21 -

Marcus Hutter

Coefficient Shrinkage

Solution 2: Shrinkage methods: Shrink the least squares w by penalizing the Loss: Ridge regression: Add ∝ ||w||2

2.

Lasso: Add ∝ ||w||1. Bayesian linear regression:

• Comp. MAP arg maxw P(w|D)

from prior P(w) and sampling model P(D|w). Weights of low variance components shrink most.

Linear Methods for Regression

• 22 -

Marcus Hutter

Linear Methods for Classification

Example: Y = {spam,non-spam} ≃ {−1, 1} (or {0, 1}) Reduction to regression: Regard y ∈ I R ⇒ ˆ w from linear regression. Binary classification: If f ˆ

w(x) > 0 then ˆ

y = 1 else ˆ y = −1. Probabilistic classification: Predict probability that new x is in class y. Log-odds log P(y=1|x,D)

P(y=0|x,D) := f ˆ w(x)

Improvements:

• Linear Discriminant Analysis (LDA)
• Logistic Regression
• Perceptron
• Maximum Margin Hyperplane
• Support Vector Machine

Generalization to non-binary Y possible.

Linear Methods for Regression

• 23 -

Marcus Hutter

Linear Basis Function Regression (LBFR)

= powerful generalization of and reduction to linear regression! Problem: Response y ∈ I R is often not linear in x ∈ I Rd. Solution: Transform x ❀ φ(x) with φ : I Rd → I Rp. Assume/hope response y ∈ I R is now linear in φ. Examples:

• Linear regression: p = d and φi(x) = xi.
• Polynomial regression: d = 1 and φi(x) = xi.
• Piecewise constant regression:

E.g. d = 1 with φi(x) = 1 for i ≤ x < i + 1 and 0 else.

• Piecewise polynomials ...
• Splines ...

Linear Methods for Regression

• 24 -

Marcus Hutter

Linear Methods for Regression

• 25 -

Marcus Hutter

2D Spline LBFR and 1D Symmlet-8 Wavelets

Linear Methods for Regression

• 26 -

Marcus Hutter

Local Smoothing & Kernel Regression

Estimate f(x) by averaging the yi for all xi a-close to x: ˆ f(x) =

Pn

i=1 K(x,xi)yi

Pn

i=1 K(x,xi)

Nearest-Neighbor Kernel: K(x, xi) = 1 if { |x−xi|<a

and 0 else

Generalization to other K, e.g. quadratic (Epanechnikov) Kernel: K(x, xi) = max{0, a2 − (x − xi)2}

Linear Methods for Regression

• 27 -

Marcus Hutter

Regularization & 1D Smoothing Splines

to avoid overfitting if function class is large ˆ f = arg minf{n

i=1(yi−f(xi))2 + λ

• (f ′′(x))2dx}
• λ = 0

⇒ ˆ f =any function through data

• λ = ∞

⇒ ˆ f=least squares line fit

• 0 < λ < ∞

⇒ ˆ f=piecwise cubic with continuous derivative

Nonlinear Regression

• 28 -

Marcus Hutter

3 NONLINEAR REGRESSION

• Artificial Neural Networks
• Kernel Trick
• Maximum Margin Classifier
• Sparse Kernel Methods / SVMs

Nonlinear Regression

• 29 -

Marcus Hutter

Artificial Neural Networks 1

as non-linear function approximator Single hidden layer feed- forward neural network

• Hidden layer: zj = h(d

i=0 w(1) ji xi)

• Output: fw,k(x) = σ(M

j=0 w(2) kj zj)

• Sigmoidal activation functions:

h() and σ() ⇒ f non-linear

• Goal: Find network weights

best modeling the data:

• Back-propagation algorithm:

Minimize n

i=1 ||yi − f w(xi)||2 2

w.r.t. w by gradient decent.

Nonlinear Regression

• 30 -

Marcus Hutter

Artificial Neural Networks 2

• Avoid overfitting by early stopping or small M.
• Avoid local minima by random initial weights
• r stochastic gradient descent.

Example: Image Processing

Nonlinear Regression

• 31 -

Marcus Hutter

Kernel Trick

The Kernel-Trick allows to reduce the functional minimization to finite-dimensional optimization.

• Let L(y, f(x)) be any loss function
• and J(f) be a penalty quadratic in f.
• then minimum of penalized loss n

i=1 L(yi, f(xi)) + λJ(f)

• has form f(x) = n

i=1 αiK(x, xi)

• with α minimizing n

i=1 L(yi, (Kα)i) + λα ⊤Kα.

• and Kernel Kij = K(xi, xj) following from J.

Nonlinear Regression

• 32 -

Marcus Hutter

Maximum Margin Classifier

ˆ w = arg max

w:||w||=1 min i {yi(w ⊤φ(xi))}

with y ∈ {−1, 1}

• Linear boundary for

φb(x) = x(b).

• Boundary is determined by

Support Vectors (circled data points)

• Margin negative if

classes not separable.

Nonlinear Regression

• 33 -

Marcus Hutter

Sparse Kernel Methods / SVMs

Non-linear boundary for general φb(x) ˆ w = n

i=1 aiφ(xi) for some a.

⇒ ˆ f(x) = ˆ w

⊤φ(x) = n i=1 aiK(xi, x)

depends only on φ via Kernel K(xi, x) = d

b=1 φb(xi)φb(x).

⇒ Huge time savings if d ≫ n Example K(x, x′):

• polynomial (1 + x, x′)d,
• Gaussian exp(−||x − x′||2

2),

• neural network tanh(x, x′).

Model Assessment & Selection

• 34 -

Marcus Hutter

4 MODEL ASSESSMENT & SELECTION

• Example: Polynomial Regression
• Training=Empirical versus Test=True Error
• Empirical Model Selection
• Theoretical Model Selection
• The Loss Rank Principle for Model Selection

Model Assessment & Selection

• 35 -

Marcus Hutter

Example: Polynomial Regression

• Straight line does not fit data well (large training error)

high bias ⇒ poor predictive performance

• High order polynomial fits data perfectly (zero training error)

high variance (overfitting) ⇒ poor prediction too!

• Reasonable polynomial

degree d performs well. How to select d? minimizing training error

• bviously does not work.

Model Assessment & Selection

• 36 -

Marcus Hutter

Training=Empirical versus Test=True Error

• Learn functional relation f for data D = {(x1, y1), ..., (xn, yn)}.
• We can compute the empirical error on past data:

ErrD(f) = 1

n

n

i=1(yi − f(xi))2.

• Assume data D is sample from some distribution P.
• We want to know the expected true error on future examples:

ErrP(f) = EP[(y − f(x))].

• How good an estimate of ErrP(f) is ErrD(f) ?
• Problem: ErrD(f) decreases with increasing model complexity,

but not ErrP(f).

Model Assessment & Selection

• 37 -

Marcus Hutter

Empirical Model Selection

How to select complexity parameter

• Kernel width a,
• penalization constant λ,
• number k of nearest neighbors,
• the polynomial degree d?

Empirical test-set-based methods: Regress on training set and mini- mize empirical error w.r.t. “com- plexity” parameter (a, λ, k, d) on a separate test-set. Sophistication: cross-validation, bootstrap, ...

Model Assessment & Selection

• 38 -

Marcus Hutter

Theoretical Model Selection

How to select complexity or flexibility or smoothness parameter: Kernel width a, penalization constant λ, number k of nearest neighbors, the polynomial degree d? For parametric regression with d parameters:

• Bayesian model selection,
• Akaike Information Criterion (AIC),
• Bayesian Information Criterion (BIC),
• Minimum Description Length (MDL),

They all add a penalty proportional to d to the loss. For non-parametric linear regression:

• Add trace of on-data regressor = effective # of parameters to loss.
• Loss Rank Principle (LoRP).

Model Assessment & Selection

• 39 -

Marcus Hutter

The Loss Rank Principle for Model Selection

Let ˆ f c

D : X → Y be the (best) regressor of complexity c on data D.

The loss Rank of ˆ f c

D is defined as the number of other (fictitious) data

D′ that are fitted better by ˆ f c

D′ than D is fitted by ˆ

f c

D.

• c is small ⇒ ˆ

f c

D fits D badly ⇒ many other D′ can be fitted better

⇒ Rank is large.

• c is large ⇒ many D′ can be fitted well ⇒ Rank is large.
• c is appropriate ⇒ ˆ

f c

D fits D well and not too many other D′

can be fitted well ⇒ Rank is small. LoRP: Select model complexity c that has minimal loss Rank Unlike most penalized maximum likelihood variants (AIC,BIC,MDL),

• LoRP only depends on the regression and the loss function.
• It works without a stochastic noise model, and
• is directly applicable to any non-parametric regressor, like kNN.

How to Attack Large Problems

• 40 -

Marcus Hutter

5 HOW TO ATTACK LARGE PROBLEMS

• Probabilistic Graphical Models (PGM)
• Trees Models
• Non-Parametric Learning
• Approximate (Variational) Inference
• Sampling Methods
• Combining Models
• Boosting

How to Attack Large Problems

• 41 -

Marcus Hutter

Probabilistic Graphical Models (PGM)

Visualize structure of model and (in)dependence ⇒ faster and more comprehensible algorithms

• Nodes = random variables.
• Edges = stochastic dependencies.
• Bayesian network = directed PGM
• Markov random field = undirected PGM

Example: P(x1)P(x2)P(x3)P(x4|x1x2x3)P(x5|x1x3)P(x6|x4)P(x7|x4x5)

How to Attack Large Problems

• 42 -

Marcus Hutter

Additive Models & Trees & Related Methods

Generalized additive model: f(x) = α + f1(x1) + ... + fd(xd) Reduces determining f : I Rd → I R to d 1d functions fb : I R → I R Classification/decision trees: Outlook:

• PRIM,
• bump hunting,
• How to learn

tree structures.

How to Attack Large Problems

• 43 -

Marcus Hutter

Regression Trees

f(x) = cb for x ∈ Rb, and cb = Average[yi|xi ∈ Rb]

How to Attack Large Problems

• 44 -

Marcus Hutter

Non-Parametric Learning

= prototype methods = instance-based learning = model-free Examples:

• K-means:

Data clusters around K centers with cluster means µ1, ...µK. Assign xi to closest cluster center.

• K Nearest neighbors regression (kNN):

Estimate f(x) by averaging the yi for the k xi closest to x:

• Kernel regression:

Take a weighted average ˆ f(x) = n

i=1 K(x, xi)yi

n

i=1 K(x, xi) .

How to Attack Large Problems

• 45 -

Marcus Hutter

How to Attack Large Problems

• 46 -

Marcus Hutter

Approximate (Variational) Inference

Approximate full distribution P(z) by q(z) Examples: Gaussians Popular: Factorized distribution: q(z) = q1(z1) × ... × qM(zM). Measure of fit: Relative entropy KL(p||q) =

• p(z) log p(z)

q(z)dz

Red curves: Left min- imizes KL(P||q), Mid- dle and Right are the two local minima of KL(q||P).

How to Attack Large Problems

• 47 -

Marcus Hutter

Elementary Sampling Methods

How to sample from P : Z → [0, 1]?

• Special sampling algorithms for standard distributions P.
• Rejection sampling: Sample z uniformly from domain Z,

but accept sample only with probability ∝ P(z).

• Importance sampling:

E[f] =

• f(z)p(z)dz ≃ 1

L

L

l=1 f(zl)p(zl)/q(zl),

where zl are sampled from q. Choose q easy to sample and large where f(zl)p(zl) is large.

• Others: Slice sampling

How to Attack Large Problems

• 48 -

Marcus Hutter

Markov Chain Monte Carlo (MCMC) Sampling

Metropolis: Choose some conve- nient q with q(z|z′) = q(z′|z). Sample zl+1 from q(·|zl) but accept

• nly

with probability min{1, p(zl+1)/p(zl)}. Gibbs Sampling: Metropolis with q leaving z unchanged from l ❀ l + 1,except resample coordinate i from P(zi|z\i). Green lines are accepted and red lines are rejected Metropolis steps.

How to Attack Large Problems

• 49 -

Marcus Hutter

Combining Models

Performance can often be improved by combining multiple models in some way, instead of just using a single model in isolation

• Committees: Average the predictions of a set of individual models.
• Bayesian model averaging: P(x) =

Models P(x|Model)P(Model)

• Mixture models: P(x|θ, π) =

k πkPk(x|θk)

• Decision tree: Each model is responsible

for a different part of the input space.

• Boosting: Train many weak classifiers in sequence

and combine output to produce a powerful committee.

How to Attack Large Problems

• 50 -

Marcus Hutter

Boosting

Idea: Train many weak classifiers Gm in sequence and combine output to produce a powerful committee G. AdaBoost.M1: [Freund & Schapire (1997) received famous G¨

• del-Prize]

Initialize observation weights wi uniformly. For m = 1 to M do: (a) Gm classifies x as Gm(x) ∈ {−1, 1}. Train Gm weighing data i with wi. (b) Give Gm high/low weight αi if it performed well/bad. (c) Increase attention=weight wi for obs. xi misclassified by Gm. Output weighted majority vote: G(x) = sign(M

m=1 αmGm(x))

Unsupervised Learning

• 51 -

Marcus Hutter

6 UNSUPERVISED LEARNING

• K-Means Clustering
• Mixture Models
• Expectation Maximization Algorithm

Unsupervised Learning

• 52 -

Marcus Hutter

Unsupervised Learning

Supervised learning: Find functional relation- ship f : X → Y from I/O data {(xi, yi)} Unsupervised learning: Find pattern in data (x1, ..., xn) without an explicit teacher (no y values). Example: Clustering e.g. by K-means Implicit goal: Find simple explanation, i.e. compress data (MDL, Occam’s razor). Density estimation: From which probability distribution P are the xi drawn?

Unsupervised Learning

• 53 -

Marcus Hutter

K-means Clustering and EM

• Data points seem to cluster around two centers.
• Assign each data point i to a cluster ki.
• Let µk= center of cluster k.
• Distortion measure: Total distance2
• f data points from cluster centers:

J(k, µ) := n

i=1 ||xi − µki||2

• Choose centers µk initially at random.
• M-step: Minimize J w.r.t. k:

Assign each point to closest center

• E-step: Minimize J w.r.t. µ:

Let µk be the mean of points belonging to cluster k

• Iteration of M-step and E-step converges to local minimum of J.

Unsupervised Learning

• 54 -

Marcus Hutter

Iterations of K-means EM Algorithm

Unsupervised Learning

• 55 -

Marcus Hutter

Mixture Models and EM

Mixture of Gaussians: P(x|πµΣ) = K

i=1 Gauss(x|µk, Σk)πk

Maximize likelihood P(x|...) w.r.t. π, µ, Σ. E-Step: Compute proba- bility γik that data point i belongs to cluster k, based

• n estimates π, µ, Σ.

M-Step: Re-estimate π, µ, Σ (take empirical mean/variance) given γik.

Non-IID: Sequential & (Re)Active Settings

• 56 -

Marcus Hutter

7 NON-IID: SEQUENTIAL & (RE)ACTIVE SETTINGS

• Sequential Data
• Sequential Decision Theory
• Learning Agents
• Reinforcement Learning (RL)
• Learning in Games

Non-IID: Sequential & (Re)Active Settings

• 57 -

Marcus Hutter

Sequential Data

General stochastic time-series: P(x1...xn) = n

i=1 P(xi|x1...xi−1)

Independent identically distributed (roulette,dice,classification): P(x1...xn) = P(x1)P(x2)...P(xn) First order Markov chain (Backgammon): P(x1...xn) = P(x1)P(x2|x1)P(x3|x2)...P(xn|xn−1) Second order Markov chain (mechanical systems): P(x1...xn) = P(x1)P(x2|x1)P(x3|x1x2)× ... × P(xn|xn−1xn−2) Hidden Markov model (speech recognition): P(x1...xn) =

• P(z1)P(z2|z1)...P(zn|zn−1)

× n

i=1 P(xi|zi)dz

Non-IID: Sequential & (Re)Active Settings

• 58 -

Marcus Hutter

Sequential Decision Theory

Setup: For t = 1, 2, 3, 4, ... Given sequence x1, x2, ..., xt−1 (1) predict/make decision yt, (2) observe xt, (3) suffer loss Loss(xt, yt), (4) t → t + 1, goto (1) Example: Weather Forecasting xt ∈ X = {sunny, rainy} yt ∈ Y = {umbrella, sunglasses} Loss sunny rainy umbrella 0.1 0.3 sunglasses 0.0 1.0 Goal: Minimize expected Loss: ˆ yt = arg minytE[Loss(xt, yt)|x1...xt−1] Greedy minimization of expected loss is optimal if: Important: Decision yt does not influence env. (future observations). Examples: Loss = square / absolute / 0-1 error function ˆ y = mean / median / mode

• f P(xt| · · · )

Non-IID: Sequential & (Re)Active Settings

• 59 -

Marcus Hutter

Learning Agents 1

? agent percepts sensors actions environment actuators

Additional complication: Learner can influence environment, and hence what he observes next. ⇒ farsightedness, planning, exploration necessary. Exploration ⇔ Exploitation problem

Non-IID: Sequential & (Re)Active Settings

• 60 -

Marcus Hutter

Learning Agents 2

Performance standard

Agent Environment

Sensors

Performance element changes knowledge learning goals Problem generator feedback Learning element Critic

Actuators

Non-IID: Sequential & (Re)Active Settings

• 61 -

Marcus Hutter

Reinforcement Learning (RL)

r1 | o1 r2 | o2 r3 | o3 r4 | o4 r5 | o5 r6 | o6 ... y1 y2 y3 y4 y5 y6 ... Agent p Environ- ment q

✟ ✟ ✟ ✟ ✙ ❍ ❍ ❍ ❍ ❨ ✏✏✏✏✏ ✏ ✶ PPPPP P q

• RL is concerned with how an agent ought to take actions

in an environment so as to maximize its long-term reward.

• Find policy that maps states of the world

to the actions the agent ought to take in those states.

• The environment is typically formulated

as a finite-state Markov decision process (MDP).

Non-IID: Sequential & (Re)Active Settings

• 62 -

Marcus Hutter

Learning in Games

• Learning though self-play.
• Backgammon (TD-Gammon).
• Samuel’s checker program.

Summary

• 63 -

Marcus Hutter

8 SUMMARY

• Important Loss Functions
• Learning Algorithm Characteristics Comparison
• More Learning
• Data Sets
• Journals
• Annual Conferences
• Recommended Literature

Summary

• 64 -

Marcus Hutter

Important Loss Functions

Summary

• 65 -

Marcus Hutter

Learning Algorithm Characteristics Comparison

Summary

• 66 -

Marcus Hutter

More Learning

• Concept Learning
• Bayesian Learning
• Computational Learning Theory (PAC learning)
• Genetic Algorithms
• Learning Sets of Rules
• Analytical Learning

Summary

• 67 -

Marcus Hutter

Data Sets

• UCI Repository:

http://www.ics.uci.edu/ mlearn/MLRepository.html

• UCI KDD Archive:

http://kdd.ics.uci.edu/summary.data.application.html

• Statlib:

http://lib.stat.cmu.edu/

• Delve:

http://www.cs.utoronto.ca/ delve/

Summary

• 68 -

Marcus Hutter

Journals

• Journal of Machine Learning Research
• Machine Learning
• IEEE Transactions on Pattern Analysis and Machine Intelligence
• Neural Computation
• Neural Networks
• IEEE Transactions on Neural Networks
• Annals of Statistics
• Journal of the American Statistical Association
• ...

Summary

• 69 -

Marcus Hutter

Annual Conferences

• Algorithmic Learning Theory (ALT)
• Computational Learning Theory (COLT)
• Uncertainty in Artificial Intelligence (UAI)
• Neural Information Processing Systems (NIPS)
• European Conference on Machine Learning (ECML)
• International Conference on Machine Learning (ICML)
• International Joint Conference on Artificial Intelligence (IJCAI)
• International Conference on Artificial Neural Networks (ICANN)

Summary

• 70 -

Marcus Hutter

Recommended Literature

[Bis06]

• C. M. Bishop. Pattern Recognition and Machine Learning.

Springer, 2006. [HTF01] T. Hastie, R. Tibshirani, and J. H. Friedman. The Elements of Statistical Learning. Springer, 2001. [Alp04]

• E. Alpaydin. Introduction to Machine Learning.

MIT Press, 2004. [Hut05]

• M. Hutter. Universal Artificial Intelligence:

Sequential Decisions based on Algorithmic Probability. Springer, Berlin, 2005. http://www.hutter1.net/ai/uaibook.htm.