added HW1 solution...

... and a few other bits

Week 1 -- Likelihoods and linear models/Homework solution/HW1.R

@@ -0,0 +1,114 @@
+
+setwd( "C:/Users/James.Thorson/Desktop/Project_git/2018_FSH556/Week 1 -- Likelihoods and linear models/Homework solution" )
+
+library(TMB)
+library( TMBdebug )  # Used to debug TMB models in windows (doesn't hurt for other operating systems)
+library( SpatialDeltaGLMM )
+library( statmod ) # Inverse-gaussian distribution pinvgauss
+
+# compile template file
+compile( "HW1.cpp" )
+dyn.load( dynlib("HW1") )
+
+###########
+# Part 1 -- analysis of real data for Alaska pollock
+###########
+
+# Settings
+set.seed(1)
+K = 10
+
+# Load data and divide into partitions
+data( EBS_pollock_data )
+CPUE = EBS_pollock_data$catch
+X = cbind( "Intercept"=rep(1,length(CPUE)) )
+Partition_i = sample( x=1:K, size=length(CPUE), replace=TRUE )
+
+# Stuff to record
+PredNLL_kj = matrix(NA, ncol=3, nrow=K)
+Table1 = array(NA, dim=c(3,3), dimnames=list(c("lognormal","gamma","invgauss"),c("negloglike","num_params","pred_negloglike_per_datum")) )
+
+# Cross-validation
+for(j in 1:3){
+for(k in 1:K){
+  Params = list("b_j"=rep(0,ncol(X)), "theta_z"=c(0,0))
+  Data = list( "Options_vec"=j-1, "y_i"=CPUE, "X_ij"=X, predTF_i=ifelse(Partition_i==k,1,0) )
+  Obj = MakeADFun( data=Data, parameters=Params, DLL="HW1")
+  Obj$env$beSilent()
+
+  # Step 3 -- test and optimize
+  Opt = TMBhelper::Optimize( obj=Obj, newtonsteps=1 )
+
+  # Extract stuff
+  Report = Obj$report()
+  PredNLL_kj[k,j] = Report$pred_jnll / sum(Partition_i==k)
+  if( PredNLL_kj[k,j]>10 ) stop("Check results")
+}}
+
+# Build table of results
+for(j in 1:3){
+  Params = list("b_j"=rep(0,ncol(X)), "theta_z"=c(0,0))
+  Data = list( "Options_vec"=j-1, "y_i"=CPUE, "X_ij"=X, predTF_i=rep(0,length(CPUE)) )
+  Obj = MakeADFun( data=Data, parameters=Params, DLL="HW1")
+
+  # Step 3 -- test and optimize
+  for(i in 1:3) Opt = TMBhelper::Optimize( obj=Obj, newtonsteps=1 )
+  Table1[j,] = c(Opt$objective, length(Opt$par), mean(PredNLL_kj[,j]))
+}
+
+# Look at results
+print(Table1)
+
+###########
+# Part 2 -- Simulation design
+###########
+
+# Settings
+set.seed(1)
+n_rep = 100
+
+# Stuff to record
+Results = array(NA, dim=c(n_rep,3,3,2), dimnames=list(NULL,paste("True=",c("lognormal","gamma","invgauss"),sep=""),paste("Est=",c("lognormal","gamma","invgauss"),sep=""),c("True","Est")) )
+
+# Loop through simulation scenarios
+for(simI in 1:3){
+
+  # Get parameter values from fitting to real data
+  Params = list("b_j"=rep(0,ncol(X)), "theta_z"=c(0,0))
+  Data = list( "Options_vec"=simI-1, "y_i"=CPUE, "X_ij"=X, predTF_i=rep(0,length(CPUE)) )
+  Obj = MakeADFun( data=Data, parameters=Params, DLL="HW1")
+  Opt_real = nlminb( start=Obj$env$last.par.best, objective=Obj$fn, gradient=Obj$gr )
+
+  # Loop through replicates and estimation models
+  for(repI in 1:n_rep){
+
+    # Simulate data
+    if(simI==1) ysim_i = rbinom(n=length(CPUE), size=1, prob=1-plogis(Opt_real$par)[2]) * rlnorm(n=length(CPUE), meanlog=Opt_real$par[1], sdlog=exp(Opt_real$par[3]))
+    if(simI==2) ysim_i = rbinom(n=length(CPUE), size=1, prob=1-plogis(Opt_real$par)[2]) * rgamma(n=length(CPUE), shape=1/(exp(Opt_real$par[3])^2), scale=exp(Opt_real$par[1])*(exp(Opt_real$par[3])^2))
+    if(simI==3) ysim_i = rbinom(n=length(CPUE), size=1, prob=1-plogis(Opt_real$par)[2]) * rinvgauss(n=length(CPUE), mean=exp(Opt_real$par[1]), shape=exp(Opt_real$par[3]))
+
+    # Loop through estimation models
+    for(estI in 1:3){
+      # Get parameter values
+      Params = list("b_j"=rep(0,ncol(X)), "theta_z"=c(0,0))
+      Data = list( "Options_vec"=estI-1, "y_i"=ysim_i, "X_ij"=X, predTF_i=rep(0,length(CPUE)) )
+      Obj = MakeADFun( data=Data, parameters=Params, DLL="HW1")
+      Opt = nlminb( start=Obj$env$last.par.best, objective=Obj$fn, gradient=Obj$gr )
+
+      # Record estimated and true intercept
+      Results[repI,simI,estI,1] = Opt_real$par[1]
+      Results[repI,simI,estI,2] = Opt$par[1]
+    }
+  }
+}
+
+# Plot
+par( mfrow=c(3,3), mar=c(3,3,0,0), mgp=c(2,0.5,0), tck=-0.02, yaxs="i")
+for( simI in 1:3 ){
+for( estI in 1:3 ){
+  hist( Results[,simI,estI,2], xlim=range(Results,na.rm=TRUE) )
+  abline( v=Results[1,simI,estI,1], lwd=3, lty="dotted" )
+}}
+
+
+

Week 1 -- Likelihoods and linear models/Homework solution/HW1.cpp

@@ -0,0 +1,69 @@
+
+#include <TMB.hpp>
+
+// dlnorm
+template<class Type>
+Type dlognorm(Type x, Type meanlog, Type sdlog, int give_log=0){
+  //return 1/(sqrt(2*M_PI)*sd)*exp(-.5*pow((x-mean)/sd,2));
+  Type logres = dnorm( log(x), meanlog, sdlog, true) - log(x);
+  if(give_log) return logres; else return exp(logres);
+}
+
+// dlnorm
+template<class Type>
+Type dinvgauss(Type x, Type mean, Type shape, int give_log=0){
+  Type logres = 0.5*log(shape) - 0.5*log(2*M_PI*pow(x,3)) - (shape * pow(x-mean,2) / (2*pow(mean,2)*x));
+  if(give_log) return logres; else return exp(logres);
+}
+
+// Space time
+template<class Type>
+Type objective_function<Type>::operator() ()
+{
+  // Options
+  DATA_IVECTOR( Options_vec );
+
+  // Data
+  DATA_VECTOR( y_i );
+  DATA_MATRIX( X_ij )
+  DATA_VECTOR( predTF_i );
+
+  // Parameters
+  PARAMETER_VECTOR( b_j );
+  PARAMETER_VECTOR( theta_z );
+
+  // Objective funcction
+  Type zero_prob = 1 / (1 + exp(-theta_z(0)));
+  Type logsd = exp(theta_z(1));
+  int n_data = y_i.size();
+  vector<Type> jnll_i(n_data);
+  jnll_i.setZero();
+  Type jnll = 0;
+  Type pred_jnll = 0;
+
+  // Linear predictor
+  vector<Type> linpred_i( n_data );
+  linpred_i = X_ij*b_j;
+
+  // Probability of data conditional on fixed effect values
+  for( int i=0; i<n_data; i++){
+    if(y_i(i)==0){
+      jnll_i(i) -= log( zero_prob );
+    }else{
+      if( Options_vec(0)==0 ) jnll_i(i) -= log( 1-zero_prob ) + dlognorm( y_i(i), linpred_i(i), logsd, true );
+      if( Options_vec(0)==1 ) jnll_i(i) -= log( 1-zero_prob ) + dgamma( y_i(i), 1/pow(logsd,2), exp(linpred_i(i))*pow(logsd,2), true );
+      if( Options_vec(0)==2 ) jnll_i(i) -= log( 1-zero_prob ) + dinvgauss( y_i(i), exp(linpred_i(i)), logsd, true );
+    }
+    // Running counter
+    if( predTF_i(i)==0 ) jnll += jnll_i(i);
+    if( predTF_i(i)==1 ) pred_jnll += jnll_i(i);
+  }
+
+  // Reporting
+  REPORT( zero_prob );
+  REPORT( logsd );
+  REPORT( linpred_i );
+  REPORT( pred_jnll );
+  REPORT( jnll_i );
+  return jnll;
+}

Week 4 -- Multivariate models/Lab/multivariate_kalman.R

@@ -112,7 +112,7 @@ compile( paste0(Version,".cpp") )
 
 # Build object
 dyn.load( dynlib("multivariate_kalman_with_simulation") )
-Obj = MakeADFun(data=Data, parameters=Parameters, random=Random, map=Map)  #
+Obj = MakeADFun(data=Data, parameters=Parameters, random=Random, map=Map, DLL="multivariate_kalman_with_simulation")  #
 
 # Optimize
 Opt = TMBhelper::Optimize( obj=Obj, getsd=TRUE, newtonsteps=1 )
@@ -122,8 +122,8 @@ if(FALSE){
   # Do simulation
   SimPar = matrix( NA, nrow=100, ncol=length(Opt$par), dimnames=list(NULL,names(Opt$par)) )
   for( i in 1:nrow(SimPar) ){
-    SimData <- Obj$simulate(par=Obj$env$last.par, complete=TRUE)
-    Obj2 <- MakeADFun(data=SimData, parameters=Parameters, random=Random, map=Map, silent=TRUE, DLL="multivariate_kalman_with_simulation")
+    SimData <- Obj$simulate( par=Obj$env$last.par, complete=TRUE )
+    Obj2 <- MakeADFun(data=SimData['y_tp'], parameters=Parameters, random=Random, map=Map, silent=TRUE, DLL="multivariate_kalman_with_simulation")
     SimPar[i,] = TMBhelper::Optimize(obj=Obj2, getsd=FALSE, newtonsteps=1)$par
     if( i%%10 == 0 ) message(paste0("Finished replicate ",i))
   }