lecture24
DESCRIPTION
TRANSCRIPT
Introduction to MachineIntroduction to Machine LearningLearning
Lecture 24Lecture 24Learning Classifier Systems
Albert Orriols i Puightt // lb t i l thttp://www.albertorriols.net
Artificial Intelligence – Machine Learningg gEnginyeria i Arquitectura La Salle
Universitat Ramon Llull
Recap of Lecture 23Michigan-style LCSg y
EnvironmentSensorialstate RewardAction
Learning Classifier 1Classifier 2
Online rule evaluator:• XCS: Q-Learning (Sutton & Barto, 1998)
Uses Widrow-Hoff delta rule
state
Any Representation:Classifier System
Classifier 2
Classifier n
Uses Widrow-Hoff delta ruley pproduction rules,
genetic programs,perceptrons,
SVMs
Rule evolution:
SVMs
GeneticAlgorithm
Rule evolution: Typically, a GA (Holland, 75; Goldberg, 89) applied on the population.
Slide 2Artificial Intelligence Machine Learning
Recap of Lecture 23Main characteristics of XCS
Population-based method
I d d t l ifiIndependent classifiers
Works under a reinforcement learning paradigm, but can be l li d t i d l i ( d t f tialso applied to supervised learning (and to function
approximation)
Cl ifi l d b ti l ithClassifiers evolved by a genetic algorithm
Slide 3Artificial Intelligence Machine Learning
Today’s Agenda
Examples of domainsAnother step toward cognitive systems
Anticipatory Classifier System
Slide 4Artificial Intelligence Machine Learning
Applications of LCSExamples of domainsp
Reinforcement learning
S i d l iSupervised learning
[Function approximation – not seen herein]
Real-life applicationsReal life applicationsData Mining
Modeling market traders
Autonomous robotics
Modeling artificial ecosystems
Slide 5
…
Artificial Intelligence Machine Learning
Example in Reinf. LearningExample: simple maze problemp p p
XCS solves more complex reinforcement learning prob :XCS solves more complex reinforcement learning prob.:Complex mazes
Slide 6
Mountain car
Artificial Intelligence Machine Learning
Example in Reinf. LearningPerformance in the Maze6 problem (Butz et al.)p ( )
Slide 7Artificial Intelligence Machine Learning
Example in Sup. LearningSolving large, non-linear, complex boolean functions
000 0#######:0000 0#######:0000 1#######:1001 #0######:0001 #1######:1001 #1######:1010 ##0#####:0010 ##1#####:1011 ###0####:0011 ###0####:0011 ###1####:1100 ####0###:0100 ####1###:1100 ####1###:1101 #####0##:0101 #####1##:1110 ######0#:0
Solving multiplexer problems up to 135 input variables
110 ######1#:1111 #######0:0111 #######1:1
Slide 8Artificial Intelligence Machine Learning
Solving multiplexer problems up to 135 input variables
Current Real-Life ApplicationsData mining
Most important application domain of LCSs
John H. HolmesJohn H. HolmesEpidemiologic study by means of LCSsHidden relationships among variables p gdiscovered by LCSs
Xavier Llorà et al.Better than Human Capability in Diagnosing Prostate Cancer Using Infrared Spectroscopic imagingg g
Many other applications and miners:GALE GAX GAssist UCS MIPSGALE, GAX, GAssist, UCS, MIPS …
See: Bull, Bernadó-Mansilla & Holmes (eds) Learning Classifier Systems in Data mining. Springer (2008)
Slide 9Artificial Intelligence Machine Learning
Current Real-Life ApplicationsModeling market tradersg
LETS project: Evolving artificial traders for successful market trading (Sonia Schulenburg et al, 2007)ad g (So a Sc u e bu g et a , 00 )
Evolutionary economics:Evolutionary economics:Create trend followersand value investors
Let them interact
Evolve a population ofstrategies
Slide 10Artificial Intelligence Machine Learning
Current Real-Life ApplicationsAutonomous Robotics
Robot shaping: Early efforts of Marco Dorigo and Marco Colombetti (1997)
Small mobile robots equipped with sensors and motorsRobots connected in real time by various sorts of modem cableRobots controlled by LCS, ICS, running on desktop computersConstant stream of positive/negative rewards (bucket brigade)
Tasks solved: Following lightsGather food and run homeHunt around for a light hidden behind and obstacle
Impressive results, high performance
Recent applications to model robotic problems performed in the University of West England
Slide 11
University of West England
Artificial Intelligence Machine Learning
Current Real-Life ApplicationsModeling Artificial Ecosystemsg y
Eden: Artificial Life environment (Jon McCromak, 2004)
Model of an environment where evolvable rule basedModel of an environment where evolvable rule-basedclassifier systems drive agent behavior.
Autonomous LCSs or agents compete for limitedAutonomous LCSs or agents compete for limitedresources.
Agents can:gMake and listen to soundsForage for foodgEncounter predatorsMate with each other
Goal: Maintain the audience in tension without fitness needing the
Slide 12
audience explicitly perform fitness selection
Artificial Intelligence Machine Learning
Toward Cognitive SystemsCognitive systemsg y
Cognitive systems are natural or artificial information processing systems, including those responsible for perception, p ocess g sys e s, c ud g ose espo s b e o pe cep o ,learning, reasoning and decision-making and for communication and action
LCS originally devised as cognitive systemsLCS originally devised as cognitive systems
A step furtherAnticipatory LCS
Slide 13Artificial Intelligence Machine Learning
Anticipatory LCSAnticipations influence cognitive systemsp g y
LCS learned:LCS learned:Conditions x actions x prediction
Anticipatory learning processes learnAnticipatory learning processes learnCondition x action x effect relations
Let’s see the Anticipatory LCS (ACS2)
Slide 14Artificial Intelligence Machine Learning
ACSENVIRONMENTENVIRONMENT
C1 A1 E1
Population [P]
Match Set [M]
A1σt σt+1
C1 – A1 – E1C2 – A2 – E2C3 – A3 – E3C4 – A1 – E4
C1 – A1 – E1C3 – A3 – E3C4 – A1 – E4
Match Set [M]
C1 – A1 – E1
Action Set [A]
C5 – A5 – E5C6 – A6 – E6C7 – A1 – E7C8 – A8 – E8
C4 – A1 – E4C6 – A6 – E6C7 – A1 – E7C9 – A9 – E9
C4 – A1 – E4C7 – A1 – E7
…C8 – A8 – E8C9 – A9 – E9
…
…
compareAnticipatory Learning Process
Slide 15Artificial Intelligence Machine Learning
Next Class
Bi i t f h t h fBig picture of what we have seen so far
New challenges in machine learning
Slide 16Artificial Intelligence Machine Learning
Introduction to MachineIntroduction to Machine LearningLearning
Lecture 24Lecture 24Learning Classifier Systems
Albert Orriols i Puightt // lb t i l thttp://www.albertorriols.net
Artificial Intelligence – Machine Learningg gEnginyeria i Arquitectura La Salle
Universitat Ramon Llull