Karteikarten
cs 580 final
Karten in diesem Stapel (48)
- 1Machine Learning HistoryGame Playing - Samuel
Neural Nets - Hebb
Evolution - Holland - 2Early Game Playing(Samuel)Learning via weight adjustment
Feature Extractors
Piece-wise linear board evaluation function
feedback
credit apportionment - 3Early History - ANNNeurons - linear threshold units
Networks - feed forward(rosenblatt)
Fully Connected(hebb) - 4Early ANN LearningSupervised - Backprop
Unsupervised - correlations - 5Earl Evolutionary ComputationDarwinian adaptation
Fogel - EP(evolve FSM)
Holland - Adaptive Systems(GAs) - 6Rebirth in 1980'sknowledge acquisition bottleneck
non-symbolic reps
adaptation - 7ML Rebirth FocusWhat's changing & how
Evironment, application
Computability, convergence, cost - 8Internal Aspects - What is changing?Weights, data structures, executable code
- 9Internal Aspects How are changes made?Learning rule:
Feedback + internalstate ==> new state - 10How to learn weights?ANN
Bayesian Inference Networks
Parameterized Models - 11Learning Rule: Weight Update ProcedureIndirect: reinforcement
Direct: Error Correction - 12External Aspects Form of FeedbackDirect: correct action
Indirect: reinforcement - 13External Aspects: Frequency of feedbackAfter each action
After a sequence of actions - 14External Aspects: Role of teacherType/quality of feedback:
Apprentice
Neutral
Adversarial - 15Example: NimLearning weights:
Reinforcement
No embedded Knowledge
No apportionment of credit - 16Example Nim: External AspectsFeedback - Indirect, after sequence of actions
- 17Learning Weights Parameter Options1 weighted board evaluation
2 weighted action sets
3 ANN weights - 18Learn StructuresDecision Tree
eg quinlan's C4.5 - 19Learning Executable CodeML Friendly programming language
LISP - 20Supervised LearningExternally controlled setting
- 21Supervised Learning - Feedback-- Correct Results
-- Reinforcement - 22Unsupervised Learning-- Open ended environment
-- basic goals(survive, capture etc)
-- environmental feedback - 23Semi-Supervised LearningStart with supervised, use open-ended environment to pose new problems and ask questions
- 24ANN - output-- pass through
-- thresholders
--squashers - 25ANN - topologiesFeed Forward
Feedback Loops
Fully Connected - 26ANN - FeedForward-- Inputs flow through network in order, no loops.
-- Backprop: iterative weight adjustment scheme
-- Gradient Descent
-- batch or sequential
--2 or 3 layers sufficient
-- width: degrees of freedom - 27Feedback Loopsunravel the loops to do backprop.
- 28Evolutionary Algorithms: Basic Elements-- individuals
-- fitness
-- birth/death cycle biased by fitness
-- inheritance - 29Evolutionary Algorithms: Properties-- Parallel, adaptive search
-- competition for resources
-- stochastic
--fitness optimization - 30Evolutionary Algorithms: Evolving Agentsgoal oriented breeder - evolve individuals with high fitness
- 31Evolutionary Algorithms: evolving agentsTeam-oriented problem solving - evolve populations with high fitness overall
- 32Evolutionary Algorithm: Mechanisms of changeMutation
Recombination - 33Statistical MLDominates ML Classification problems
Leading method: SVM - 34Reinforcement Learning-- animal learning techniques
-- feedback is reward/punishment
-- internal state space representation of environment
--learn S=>R rules by exploration and feedback - 35Dominant ML TasksConcept Learning
Classification - 36Other ML TasksPredictive Tasks (REgression, Time Series)
Discovery Tasks
Sequential Decision Tasks - 37Concept Learning– Given an instance space X
– Given a concept c defined on X
– Given a hypothesis space H
– Given a training set T (examples of c)
– Find the best h in H approximation of c - 38Concept Learning Exampleshand writing recog
satellite image classification
spam fitlers - 39PAC AnalysisCharacterize degree of difficult of concept learning in terms of
size/complexity of hyp. space
accuracy of produced hyp
ability to generate accurate hyp - 40Probably Approximately CorrectUsing a reasonable training set and reasonable computation time,
with probability 1-δ a learner L will produce an h with error < ε - 41Support Vector Machines– embed non-linear classification problems in
a higher dimensional space.
– “right” embedding is now linearly separable - 42Concept Learning issues-- noisy data
-- incorrect data
-- poorly chosen features
-- poorly chosen hyp spaces - 43Unsupervised Concept Learning Clusteringidentify subsets with similar properties
K-Means - 44SEquential Decision Task ExamplesGame Playing
Navigation
Pole Balancing - 45Sequential Decision Tasks - Typical PropertiesTemporal Extent
Interaction with environment
Intermittent reinforcement feedback - 46Reinforcement Learning TechniquesQ learning
TD Learning - 47Multi-agent learning How to learn?Separate, independent learners?
– E.g., each a reinforcement learner
• Single master plan?
– E.g., command structure
• Distributed, local interactions?
– E.g., ant colony algorithms - 48Batch Learning AlgsInductive generalization:
• Seed growing: extend against
• Divide & conquer: information theory
• Statistical: maximum likelihood
• Minimal descriptions (Occam’s razor)
Informationen
- Beschreibung
- Karteikarten am cs 580 final, erstellt von Jonathan Zinger am 08/05/2014.
- Erstellt von
- Jonathan Zinger
- Aufrufe
- 21
- Bewertung