Introduction We present a package for the analysis of (non-hierarchical) Multinomial Processing Tree (MPT) models for the statistical programming language.

Slides:



Advertisements
Similar presentations
Fast Algorithms For Hierarchical Range Histogram Constructions
Advertisements

Quentin Frederik Gronau*, Axel Rosenbruch, Paul Bacher, Henrik Singmann, and David Kellen Poster presented at MathPsych, Québec (2014) Validating Recognition.
CSCI 347 / CS 4206: Data Mining Module 07: Implementations Topic 03: Linear Models.
The Impact of Criterion Noise in Signal Detection Theory: An Evaluation across Recognition Memory Tasks Julie Linzer David Kellen Henrik Singmann Karl.
Chap 8: Estimation of parameters & Fitting of Probability Distributions Section 6.1: INTRODUCTION Unknown parameter(s) values must be estimated before.
ECE 8443 – Pattern Recognition ECE 8527 – Introduction to Machine Learning and Pattern Recognition Objectives: Jensen’s Inequality (Special Case) EM Theorem.
Basis Expansion and Regularization Presenter: Hongliang Fei Brian Quanz Brian Quanz Date: July 03, 2008.
Introduction to Bioinformatics
Jun Zhu Dept. of Comp. Sci. & Tech., Tsinghua University This work was done when I was a visiting researcher at CMU. Joint.
FALL 2006CENG 351 Data Management and File Structures1 External Sorting.
Elementary hypothesis testing Purpose of hypothesis testing Type of hypotheses Type of errors Critical regions Significant levels Hypothesis vs intervals.
1 Introduction: syntax and semantics Syntax: a formal description of the structure of programs in a given language. Semantics: a formal description of.
Lecture 10 Comparison and Evaluation of Alternative System Designs.
BCOR 1020 Business Statistics
Regression Eric Feigelson. Classical regression model ``The expectation (mean) of the dependent (response) variable Y for a given value of the independent.
M obile C omputing G roup A quick-and-dirty tutorial on the chi2 test for goodness-of-fit testing.
SENG521 (Fall SENG 521 Software Reliability & Testing Software Reliability Tools (Part 8a) Department of Electrical & Computer.
The BioAnalytics Group LLC Global Optimization Toolkit Project First Prototype Delivery.
Introduction to MCMC and BUGS. Computational problems More parameters -> even more parameter combinations Exact computation and grid approximation become.
Kansas State University Department of Computing and Information Sciences CIS 736: Computer Graphics Wednesday, February 23, 2000 William H. Hsu Department.
1 Experimental Statistics - week 4 Chapter 8: 1-factor ANOVA models Using SAS.
One-class Training for Masquerade Detection Ke Wang, Sal Stolfo Columbia University Computer Science IDS Lab.
Measurement Models: Identification and Estimation James G. Anderson, Ph.D. Purdue University.
Quentin Frederik Gronau, Axel Rosenbruch, Paul Bacher, Henrik Singmann, David Kellen Poster presented at the TeaP, Gießen (2014) Validating a Two-High.
Lecture 3: Statistics Review I Date: 9/3/02  Distributions  Likelihood  Hypothesis tests.
Topic Models Presented by Iulian Pruteanu Friday, July 28 th, 2006.
Cluster validation Integration ICES Bioinformatics.
School of Computer Science 1 Information Extraction with HMM Structures Learned by Stochastic Optimization Dayne Freitag and Andrew McCallum Presented.
ECE 8443 – Pattern Recognition Objectives: Jensen’s Inequality (Special Case) EM Theorem Proof EM Example – Missing Data Intro to Hidden Markov Models.
- 1 - Calibration with discrepancy Major references –Calibration lecture is not in the book. –Kennedy, Marc C., and Anthony O'Hagan. "Bayesian calibration.
Digital Design Using VHDL and PLDs ECOM 4311 Digital System Design Chapter 1.
Chapter 2-OPTIMIZATION G.Anuradha. Contents Derivative-based Optimization –Descent Methods –The Method of Steepest Descent –Classical Newton’s Method.
ALISON BOWLING MAXIMUM LIKELIHOOD. GENERAL LINEAR MODEL.
1 Chapter 8: Model Inference and Averaging Presented by Hui Fang.
A Brief Maximum Entropy Tutorial Presenter: Davidson Date: 2009/02/04 Original Author: Adam Berger, 1996/07/05
Asymptotic Behavior of Stochastic Complexity of Complete Bipartite Graph-Type Boltzmann Machines Yu Nishiyama and Sumio Watanabe Tokyo Institute of Technology,
Anders Nielsen Technical University of Denmark, DTU-Aqua Mark Maunder Inter-American Tropical Tuna Commission An Introduction.
Hypothesis Testing. Statistical Inference – dealing with parameter and model uncertainty  Confidence Intervals (credible intervals)  Hypothesis Tests.
Maximum Likelihood Estimates and the EM Algorithms III Henry Horng-Shing Lu Institute of Statistics National Chiao Tung University
Computational Intelligence: Methods and Applications Lecture 26 Density estimation, Expectation Maximization. Włodzisław Duch Dept. of Informatics, UMK.
Approximation Algorithms based on linear programming.
Bootstrapping James G. Anderson, Ph.D. Purdue University.
1 Minimum Bayes-risk Methods in Automatic Speech Recognition Vaibhava Geol And William Byrne IBM ; Johns Hopkins University 2003 by CRC Press LLC 2005/4/26.
Bayesian Extension to the Language Model for Ad Hoc Information Retrieval Hugo Zaragoza, Djoerd Hiemstra, Michael Tipping Microsoft Research Cambridge,
SECTION 6 DESIGN STUDY. What’s in this section: –Design Variables –Design Studies Overview –Specifying an Objective –Execution Display Settings –Output.
Ranking: Compare, Don’t Score Ammar Ammar, Devavrat Shah (LIDS – MIT) Poster ( No preprint), WIDS 2011.
Introduction toData structures and Algorithms
Requirements Specification
Combining Models Foundations of Algorithms and Machine Learning (CS60020), IIT KGP, 2017: Indrajit Bhattacharya.
Algorithms II Software Development Life-Cycle.
Unified Modeling Language
CJT 765: Structural Equation Modeling
Multimodal Learning with Deep Boltzmann Machines
Clustering (3) Center-based algorithms Fuzzy k-means
Henrik Singmann Karl Christoph Klauer Sieghard Beller
Front End vs Back End of a Compilers
Henrik Singmann David Kellen Karl Christoph Klauer
Predictive distributions
Parameter Redundancy and Identifiability in Ecological Models
Software Metrics “How do we measure the software?”
Ch13 Empirical Methods.
Chap. 7 Regularization for Deep Learning (7.8~7.12 )
David Kellen, Henrik Singmann, Sharon Chen, and Samuel Winiger
A QUICK START TO OPL IBM ILOG OPL V6.3 > Starting Kit >
Applications of Genetic Algorithms TJHSST Computer Systems Lab
CENG 351 Data Management and File Structures
Henrik Singmann (University of Warwick)
Chapter 4 . Trajectory planning and Inverse kinematics
New Results from the Bayesian and Frequentist MPT Multiverse
Fractional-Random-Weight Bootstrap
Presentation transcript:

Introduction We present a package for the analysis of (non-hierarchical) Multinomial Processing Tree (MPT) models for the statistical programming language R: MPTinR. MPTinR combines two approaches, extending the functionality of other software for MPTs: ease of use and amount of functionalities. The homepage of MPTinR (including documentation) is: Introduction We present a package for the analysis of (non-hierarchical) Multinomial Processing Tree (MPT) models for the statistical programming language R: MPTinR. MPTinR combines two approaches, extending the functionality of other software for MPTs: ease of use and amount of functionalities. The homepage of MPTinR (including documentation) is: Features The advantages of MPTinR over similar programs for fitting MPTs are: MPTinR is an R package and therefore integrates smoothly with a R-workflow. MPTinR is free software and open source! MPTinR allows an easy and intuitive way to specify the model in a file that even allows comments (via “#”). Furthermore, MPTinR supports the 'classical' EQN syntax (as GPT, HMMTree, MultiTree). Sequential equality and inequality model restrictions can be specified in an external file. MPTinR provides different outputs for individual and datasets consisting of multiple individuals. In the latter case, results for each individual, sums of the individual results, and results from aggregating the data are automatically provided. For model selection, the Fisher information approximation (FIA), a minimum description length based measure of model complexity, can be obtained using the algorithm provided by Wu, Myung and Batchelder (2010). For multiple individuals or multiple fitting runs the package allows one to easily use multiple cores (or CPUs) via the snowfall package by simply specifying the number of available cores. Features The advantages of MPTinR over similar programs for fitting MPTs are: MPTinR is an R package and therefore integrates smoothly with a R-workflow. MPTinR is free software and open source! MPTinR allows an easy and intuitive way to specify the model in a file that even allows comments (via “#”). Furthermore, MPTinR supports the 'classical' EQN syntax (as GPT, HMMTree, MultiTree). Sequential equality and inequality model restrictions can be specified in an external file. MPTinR provides different outputs for individual and datasets consisting of multiple individuals. In the latter case, results for each individual, sums of the individual results, and results from aggregating the data are automatically provided. For model selection, the Fisher information approximation (FIA), a minimum description length based measure of model complexity, can be obtained using the algorithm provided by Wu, Myung and Batchelder (2010). For multiple individuals or multiple fitting runs the package allows one to easily use multiple cores (or CPUs) via the snowfall package by simply specifying the number of available cores. Installation MPTinR is hosted on R-Forge: To install simply type at an R prompt: install.packages("MPTinR", repos=" Forge.R-project.org") Installation MPTinR is hosted on R-Forge: To install simply type at an R prompt: install.packages("MPTinR", repos=" Forge.R-project.org") * Contact: * Contact: Model Files MPTinR allows two ways two specify the model file: 1.The new “easy” format 2.The classical EQN format. Furthermore, MPTinR can create EQN model files from models specified in the “easy” format via the function make.eqn(). Next, we compare model files for the sample 2HTM model (on the right) for both formats: Model Files MPTinR allows two ways two specify the model file: 1.The new “easy” format 2.The classical EQN format. Furthermore, MPTinR can create EQN model files from models specified in the “easy” format via the function make.eqn(). Next, we compare model files for the sample 2HTM model (on the right) for both formats: Easy FormatEQN Format Trees are separated by at least one empty line. Model files can contain comments (via “#”) Needs a special syntax (see e.g., Stahl & Klauer, 2007) This file was constructed using make.eqn() #Tree for old items Do + (1 - Do) * g (1-Do)*(1-g) #Tree for new items (1-Dn) * g Dn + (1-Dn) * (1 - g) Do 1 1 (1-Do)*g 1 2 (1-Do)*(1-g) 2 4 Dn 2 3 (1-Dn)*g 2 4 (1-Dn)*(1-g) MPTinR allows two ways two specify the model file: Restrictions Files Model restrictions can be specified in an external file. MPTinR allows for sequential equality (“=“) and inequality (“<“) restrictions. Restriction files can contain comments (“#”) Models are reparametrized to accommodate the restrictions (using “method A”; Knapp & Batchelder, 2004). Restrictions Files Model restrictions can be specified in an external file. MPTinR allows for sequential equality (“=“) and inequality (“<“) restrictions. Restriction files can contain comments (“#”) Models are reparametrized to accommodate the restrictions (using “method A”; Knapp & Batchelder, 2004). # Example: #Ds are ordered D1 < D2 < D3 #D4 equal to D2 D4 = D2 # Bs set to.33 B1 = B3 = X4 = X5 = G6 Fitting Algorithm Fitting is done by the general purpose optimization routine L-BFGS-B (Byrd et al., 1995), a quasi-Newton (i.e., gradient-based) method. Multiple fitting runs (with random starting values) can be initiated to avoid local minima. Alternatively, starting values can be specified. Currently other algorithms (e.g., simplex) are being implemented. Fitting Algorithm Fitting is done by the general purpose optimization routine L-BFGS-B (Byrd et al., 1995), a quasi-Newton (i.e., gradient-based) method. Multiple fitting runs (with random starting values) can be initiated to avoid local minima. Alternatively, starting values can be specified. Currently other algorithms (e.g., simplex) are being implemented. Example: Model Fitting Fitting the 2HTM to the data from Bröder & Schütz (2009, Exp. 3) : > fit.mpt(d.broeder, m.2htm, fia = ) $goodness.of.fit Log.Likelihood G.Squared df p.value $information.criteria FIA AIC BIC $model.info rank.hessian n.parameters n.independent.categories $parameters estimates lower.conf upper.conf Dn Do G G G G G $data $data$observed [,1] [,2] [,3] [,4] [,5] [,6] [,7] [,8] [,9] [1,] [...] $data$predicted [,1] [,2] [,3] [,4] [,5] [1,] [...] For individual data the output would contain the single elements (e.g., goodness.of.fit, parameters ) for the individual data, the sum of the individual data, and the aggregated data. Example: Model Fitting Fitting the 2HTM to the data from Bröder & Schütz (2009, Exp. 3) : > fit.mpt(d.broeder, m.2htm, fia = ) $goodness.of.fit Log.Likelihood G.Squared df p.value $information.criteria FIA AIC BIC $model.info rank.hessian n.parameters n.independent.categories $parameters estimates lower.conf upper.conf Dn Do G G G G G $data $data$observed [,1] [,2] [,3] [,4] [,5] [,6] [,7] [,8] [,9] [1,] [...] $data$predicted [,1] [,2] [,3] [,4] [,5] [1,] [...] For individual data the output would contain the single elements (e.g., goodness.of.fit, parameters ) for the individual data, the sum of the individual data, and the aggregated data. Fisher Information Approximation The Fisher Information Approximation (FIA) is a minimum description length based measure of model complexity. The FIA measures complexity not only by the number of parameters (as do AIC and BIC) but estimates the functional form of the model. MPTinR can compute the FIA for any MPT using the MCMC algorithm by Wu, Myung and Batchelder (2010) ported to R. To this end MPTinR first transforms the model into the context free language of MPT Models (L-BMPT; Purdy & Batchelder, 2009). The example 2HTM (see section “Model Files”) in L-BMPT is: hj1 Do 1 g 1 2 Dn 4 g 3 4 (numbers refer to response categories, the parameter hj1 joins both trees into a single tree) Fisher Information Approximation The Fisher Information Approximation (FIA) is a minimum description length based measure of model complexity. The FIA measures complexity not only by the number of parameters (as do AIC and BIC) but estimates the functional form of the model. MPTinR can compute the FIA for any MPT using the MCMC algorithm by Wu, Myung and Batchelder (2010) ported to R. To this end MPTinR first transforms the model into the context free language of MPT Models (L-BMPT; Purdy & Batchelder, 2009). The example 2HTM (see section “Model Files”) in L-BMPT is: hj1 Do 1 g 1 2 Dn 4 g 3 4 (numbers refer to response categories, the parameter hj1 joins both trees into a single tree) Future Developments Parametric/Nonparametric Bootstrap and Jackknife methods Latent Class hierarchical modeling Normalized Maximum Likelihood approximation Increase Computing Performance Future Developments Parametric/Nonparametric Bootstrap and Jackknife methods Latent Class hierarchical modeling Normalized Maximum Likelihood approximation Increase Computing Performance References Bröder, A., & Schütz, J. (2009). Recognition ROCs are curvilinear—or are they? On premature arguments against the two-high-threshold model of recognition. Journal of Experimental Psychology: Learning, Memory, and Cognition, 35(3), 587. Byrd, R. H., Lu, P., Nocedal, J., & Zhu, C. (1995). A limited memory algorithm for bound constrained optimization. SIAM J. Scientific Computing, 16, Knapp, B. R., & Batchelder, W. H. (2004). Representing parametric order constraints in multi- trial applications of multinomial processing tree models. Journal of Mathematical Psychology, 48, Purdy, B. P., & Batchelder, W. H. (2009). A context-free language for binary multinomial processing tree models. Journal of Mathematical Psychology, 53, Stahl, C., & Klauer, K. C. (2007). HMMTree: A computer program for latent-class hierarchical multinomial processing tree models. Behavior Research Methods, 39, 267–273. Wu, H., Myung, J.I., & Batchelder, W.H. (2010). Minimum description length model selection of multinomial processing tree models. Psychonomic Bulletin & Review, 17, References Bröder, A., & Schütz, J. (2009). Recognition ROCs are curvilinear—or are they? On premature arguments against the two-high-threshold model of recognition. Journal of Experimental Psychology: Learning, Memory, and Cognition, 35(3), 587. Byrd, R. H., Lu, P., Nocedal, J., & Zhu, C. (1995). A limited memory algorithm for bound constrained optimization. SIAM J. Scientific Computing, 16, Knapp, B. R., & Batchelder, W. H. (2004). Representing parametric order constraints in multi- trial applications of multinomial processing tree models. Journal of Mathematical Psychology, 48, Purdy, B. P., & Batchelder, W. H. (2009). A context-free language for binary multinomial processing tree models. Journal of Mathematical Psychology, 53, Stahl, C., & Klauer, K. C. (2007). HMMTree: A computer program for latent-class hierarchical multinomial processing tree models. Behavior Research Methods, 39, 267–273. Wu, H., Myung, J.I., & Batchelder, W.H. (2010). Minimum description length model selection of multinomial processing tree models. Psychonomic Bulletin & Review, 17, Example: Model Selection Output of the model selection function comparing the fits of 40 individual data sets from Bröder & Schütz (2009, Exp. 3) to the original 2HTM, a restricted 2HTM (Do = Dn), and a 1HTM (Dn = 0). select.mpt() takes the outputs from fit.mpt() and creates the following table: > select.mpt(list(orig.2htm = br.2htm, res.2htm = br.2htm.res, orig.1htm = br.1htm)) model n.parameters delta.FIA.sum FIA.best delta.AIC.sum wAIC.sum AIC.best delta.BIC.sum wBIC.sum BIC.best 1 orig.2htm res.2htm orig.1htm Example: Model Selection Output of the model selection function comparing the fits of 40 individual data sets from Bröder & Schütz (2009, Exp. 3) to the original 2HTM, a restricted 2HTM (Do = Dn), and a 1HTM (Dn = 0). select.mpt() takes the outputs from fit.mpt() and creates the following table: > select.mpt(list(orig.2htm = br.2htm, res.2htm = br.2htm.res, orig.1htm = br.1htm)) model n.parameters delta.FIA.sum FIA.best delta.AIC.sum wAIC.sum AIC.best delta.BIC.sum wBIC.sum BIC.best 1 orig.2htm res.2htm orig.1htm Poster presented at the 44th Annual Meeting of the Society for Mathematical Psychology, Boston, Massachusetts, July 2011 MPTinR: An (almost) complete R package for analyzing Multinomial Processing Tree Models Henrik Singmann*, David Kellen, Fabian Hölzenbein, & Christoph Klauer University of Freiburg