Skip your coffee break: Speed up MCMC convergence

class: center, middle, title-slide

# Skip your coffee break: Speed up MCMC convergence
### The team
### last updated: 2021-05-12

---

## Our `nimble` workflow so far

![](img/nimble_workflow_sofar.png)

---
## But `nimble` gives full access to the MCMC engine

![](img/nimble_workflow.png)

---
## Steps to use NIMBLE at full capacity

1. Build the model. It is an R object.
2. Build the MCMC.
3. Compile the model and MCMC.
4. Run the MCMC.
5. Extract the samples.

- `nimbleMCMC` does all of this at once.

---
class: middle, center

# Back to CJS models with Dipper data.

---
### Define model

.tiny-font[

```r
hmm.phip <- nimbleCode({
  delta[1] <- 1              # Pr(alive t = 1) = 1
  delta[2] <- 0              # Pr(dead t = 1) = 0
    phi ~ dunif(0, 1)     # prior survival
    gamma[1,1] <- phi        # Pr(alive t -> alive t+1)
    gamma[1,2] <- 1 - phi    # Pr(alive t -> dead t+1)
    gamma[2,1] <- 0          # Pr(dead t -> alive t+1)
    gamma[2,2] <- 1          # Pr(dead t -> dead t+1)
    p ~ dunif(0, 1)       # prior detection
    omega[1,1] <- 1 - p    # Pr(alive t -> non-detected t)
    omega[1,2] <- p        # Pr(alive t -> detected t)
    omega[2,1] <- 1        # Pr(dead t -> non-detected t)
    omega[2,2] <- 0        # Pr(dead t -> detected t)
  # likelihood
  for (i in 1:N){
    z[i,first[i]] ~ dcat(delta[1:2])
    for (j in (first[i]+1):T){
      z[i,j] ~ dcat(gamma[z[i,j-1], 1:2])
      y[i,j] ~ dcat(omega[z[i,j], 1:2])
    }
  }
})
```
]

---
### Run and summarise

.tiny-font[

```r
mcmc.phip <- nimbleMCMC(code = hmm.phip, 
                         constants = my.constants,
                         data = my.data,              
                         inits = initial.values,
                         monitors = parameters.to.save,
                         niter = n.iter,
                         nburnin = n.burnin, 
                         nchains = n.chains)
```

```
|-------------|-------------|-------------|-------------|
|-------------------------------------------------------|
|-------------|-------------|-------------|-------------|
|-------------------------------------------------------|
```
]

.tiny-font[

```r
MCMCsummary(object = mcmc.phip, round = 2)
```

```
    mean   sd 2.5%  50% 97.5% Rhat n.eff
p   0.90 0.03 0.83 0.90  0.95 1.00   286
phi 0.56 0.02 0.51 0.56  0.61 1.02   541
```
]

---
class: middle, center

# Detailed Nimble workflow

---
## 1. Build the model (R object)

.tiny-font[

```r
hmm.phip <- nimbleModel(code = hmm.phip,
                        constants = my.constants,
                        data = my.data,
                        inits = initial.values())
```

```
defining model...
```

```
building model...
```

```
setting data and initial values...
```

```
running calculate on model (any error reports that follow may simply reflect missing values in model variables) ... 
checking model sizes and dimensions...
model building finished.
```
]

---
## 2. Build the MCMC

.small-font[

```r
phip.mcmc.configuration <- configureMCMC(hmm.phip)
```

```
===== Monitors =====
thin = 1: phi, p, z
===== Samplers =====
RW sampler (2)
  - phi
  - p
posterior_predictive sampler (39)
  - z[]  (39 elements)
categorical sampler (1103)
  - z[]  (1103 elements)
```

```r
phip.mcmc <- buildMCMC(phip.mcmc.configuration)
```
]

---
## 3. Compile the model and MCMC

.small-font[

```r
phip.model <- compileNimble(hmm.phip) 
```

```
compiling... this may take a minute. Use 'showCompilerOutput = TRUE' to see C++ compilation details.
```

```
compilation finished.
```

```r
c.phip.mcmc <- compileNimble(phip.mcmc, project = phip.model)
```

```
compiling... this may take a minute. Use 'showCompilerOutput = TRUE' to see C++ compilation details.
compilation finished.
```
]

---
## 4. Run the MCMC

.small-font[

```r
samples <- runMCMC(c.phip.mcmc, niter = 1000)
```

```
running chain 1...
```

```
|-------------|-------------|-------------|-------------|
|-------------------------------------------------------|
```

```r
# Alternative:
# c.phip.mcmc$run(1000)
# samples <- as.matrix(c.phip.mcmc$mvSamples)
```
]

---
## 5. Look at results

.small-font[

```r
summary(samples[,"phi"])
```

```
   Min. 1st Qu.  Median    Mean 3rd Qu.    Max. 
 0.4236  0.5472  0.5687  0.5698  0.5795  0.7440 
```

```r
summary(samples[,"p"])
```

```
   Min. 1st Qu.  Median    Mean 3rd Qu.    Max. 
 0.5468  0.8803  0.8961  0.8771  0.9128  0.9686 
```
]

---
class: middle, center

# Why is it useful?

---
## Use and debug model in `R`

+ Makes your life easier when it comes to debugging

+ Inspect variables

```r
hmm.phip$gamma
```

```
          [,1]      [,2]
[1,] 0.4235601 0.5764399
[2,] 0.0000000 1.0000000
```

+ Calculate likelihood

```r
hmm.phip$calculate()
```

```
[1] -1406.893
```

---
## Example of debugging a model in `R`

+ Pretend an impossible state was given in inits, making a dead bird alive again.

.small-font[

```r
phip.model$calculate("z")   # We can see there is a problem in z (states).
```

```
[1] -Inf
```

```r
c(phip.model$calculate("z[5,]"), # Bird 5 is valid.
  phip.model$calculate("z[6,]")) # Bird 6 isn't.
```

```
[1] -3.883928      -Inf
```

```r
phip.model$z[6,]            # We have found the problem
```

```
[1] 1 1 2 1 2 2 2
```
]

---
## Open the hood, and change/modify/write samplers

+ Slice samplers instead of Metropolis-Hastings.

+ Samplers on a log scale, especially for a variance, standard deviation, or precision parameter.

+ Blocking correlated parameters.

+ To know all samplers available in Nimble, type in `help(samplers)`.

+ Source code for samplers and distributions is **in R** and can be copied and modified.

+ Use [`compareMCMCs` package](https://github.com/nimble-dev/compareMCMCs) to compare options (including Stan and Jags!).

---
class: middle, center

## Consider a model with wing length and individual random effect on survival.

---

.tiny-font[

```r
hmm.phiwlrep <- nimbleCode({
    p ~ dunif(0, 1) # prior detection
    omega[1,1] <- 1 - p    # Pr(alive t -> non-detected t)
    omega[1,2] <- p        # Pr(alive t -> detected t)
    omega[2,1] <- 1        # Pr(dead t -> non-detected t)
    omega[2,2] <- 0        # Pr(dead t -> detected t)
  for (i in 1:N){
*   logit(phi[i]) <- beta[1] + beta[2] * winglength[i] + eps[i]
*   eps[i] ~ dnorm(mean = 0, sd = sdeps)
    gamma[1,1,i] <- phi[i]      # Pr(alive t -> alive t+1)
    gamma[1,2,i] <- 1 - phi[i]  # Pr(alive t -> dead t+1)
    gamma[2,1,i] <- 0           # Pr(dead t -> alive t+1)
    gamma[2,2,i] <- 1           # Pr(dead t -> dead t+1)
  }
  beta[1] ~ dnorm(mean = 0, sd = 1.5)
  beta[2] ~ dnorm(mean = 0, sd = 1.5)
  sdeps ~ dunif(0, 10)
  delta[1] <- 1          # Pr(alive t = 1) = 1
  delta[2] <- 0          # Pr(dead t = 1) = 0
  # likelihood
  for (i in 1:N){
    z[i,first[i]] ~ dcat(delta[1:2])
    for (j in (first[i]+1):T){
*     z[i,j] ~ dcat(gamma[z[i,j-1], 1:2, i])
      y[i,j] ~ dcat(omega[z[i,j], 1:2])
    }
  }
})
```
]

---
## Trace plot for standard deviation of the random effect (default sampler)

---
## Change samplers

+ Good sampling strategies depend on the model and data.  What are the samplers used by default?

.tiny-font[

```r
mcmcConf <- configureMCMC(hmm.phiwlrep.m)
```

```
===== Monitors =====
thin = 1: p, beta, sdeps, z
===== Samplers =====
RW sampler (259)
  - p
  - beta[]  (2 elements)
  - sdeps
  - eps[]  (255 elements)
posterior_predictive sampler (78)
  - eps[]  (39 elements)
  - z[]  (39 elements)
categorical sampler (1103)
  - z[]  (1103 elements)
```
]

---
## Remove default sampler, and use slice sampler

.tiny-font[

```r
mcmcConf$removeSamplers('sdeps')
mcmcConf$addSampler(target = 'sdeps',
*                   type = "slice")
mcmcConf
```

```
===== Monitors =====
thin = 1: p, beta, sdeps, z
===== Samplers =====
slice sampler (1)
  - sdeps
RW sampler (258)
  - p
  - beta[]  (2 elements)
  - eps[]  (255 elements)
posterior_predictive sampler (78)
  - eps[]  (39 elements)
  - z[]  (39 elements)
categorical sampler (1103)
  - z[]  (1103 elements)
```
]

---
## Trace plot for standard deviation of the random effect (slice sampler)

---
## Which is better?

+ MCMC efficiency depends on both mixing and computation time.

+ MCMC efficiency = Effective Sample Size (ESS) / computation time.

+ MCMC efficiency is the number of effectively independent posterior samples generated per second.

+ ESS is different for each parameter.  (Computation time is the same for each parameter.)

+ ESS can be estimated from packages `coda` or `mcmcse`. These give statistical estimates, so different runs will give different estimates.

+ Efficiency with default sampler = 25.7 / 21.53 = 1.19.

+ Efficiency with slice sampler = 19.24 / 19.39 = 0.99.

---
## Block sampling

+ High correlation in (regression) parameters may make independent samplers inefficient.

+ Block sampling (propose candidate values from multivariate distribution) might help.

---
## Block sampling

+ Remove and replace independent RW samples by block sampling. Then proceed as usual.

.tiny-font[

```r
mcmcConf$removeSamplers(c('beta[1]','beta[2]'))
mcmcConf$addSampler(target = c('beta[1]','beta[2]'),
*                   type = "RW_block")
```
]

---
## Block sampling

.tiny-font[

```r
mcmcConf
```

```
===== Monitors =====
thin = 1: p, beta, sdeps, z
===== Samplers =====
slice sampler (1)
  - sdeps
RW_block sampler (1)
  - beta[1], beta[2] 
RW sampler (256)
  - p
  - eps[]  (255 elements)
posterior_predictive sampler (78)
  - eps[]  (39 elements)
  - z[]  (39 elements)
categorical sampler (1103)
  - z[]  (1103 elements)
```
]

---
## Summary of strategies for improving MCMC

+ Choose better initial values.

+ Customize sampler choice (more in [Chapter 7 of the User's manual](https://r-nimble.org/html_manual/cha-mcmc.html)).

+ Reparameterize, e.g. standardize covariates, deal with parameter redundancy.

+ Rewrite the model.

+ Vectorize to improve computational efficiency (not covered).
    + Avoid long chains of deterministic dependencies.
    + Marginalize to remove parameters
    + Use new functions and new distributions written as nimbleFunctions.

+ Write new samplers that take advantage of particular model structures (not covered).

+ Using multiple cores with parallelization: see how-to at <https://r-nimble.org/nimbleExamples/parallelizing_NIMBLE.html>

---
## Marginalization

+ User-defined distributions is another neat feature of Nimble.

+ Integrate over latent states if those are not the focus of ecological inference (marginalization).

+ Marginalization often (but not always) improves MCMC. See [Ponisio et al. 2020](https://onlinelibrary.wiley.com/doi/full/10.1002/ece3.6053) for examples.

+ The [nimbleEcology](https://cran.r-project.org/web/packages/nimbleEcology/vignettes/Introduction_to_nimbleEcology.html) package implements capture-recapture models and HMMs with marginalization.

---
### Our model `$(\phi_A, \phi_B, \psi_{AB}, \psi_{BA}, p_A, p_B)$`

.small-font[

```r
multisite <- nimbleCode({
...
  # Likelihood 
  for (i in 1:N){
    # Define latent state at first capture
    z[i,first[i]] <- y[i,first[i]] - 1
    for (t in (first[i]+1):K){
      # State process: draw S(t) given S(t-1)
      z[i,t] ~ dcat(gamma[z[i,t-1],1:3])
      # Observation process: draw O(t) given S(t)
      y[i,t] ~ dcat(omega[z[i,t],1:3])
    }
  }
})
```
]

---
### Same model with nimbleEcology

.tiny-font[

```r
multisite <- nimbleCode({
...
# initial state probs
for(i in 1:N) {
  init[i, 1:4] <- gamma[ y[i, first[i] ] - 1, 1:4 ] # first state propagation
}
    
# likelihood 
for (i in 1:N){
  y[i,(first[i]+1):K] ~ dHMM(init = init[i,1:4],           # count data from first[i] + 1
                             probObs = omega[1:4,1:4],     # observation matrix
                             probTrans = gamma[1:4,1:4],   # transition matrix
                             len = K - first[i],           # nb of occasions
                             checkRowSums = 0)             # do not check whether elements in a row sum tp 1
}
...
```
]

+ This runs twice as fast as the standard formulation with explicit latent states.

+ Marginalizing typically gives better mixing.

---
### Reducing redundant calculations

+ So far, a row of the dataset is an individual. However, several individuals may share the same encounter history.

+ The contribution of `$M$` individuals with the same encounter history is the likelihood of this particular encounter history raised to the power `$M$`.

+ Using this so-called **weighted likelihood** greatly decreases the computational burden.

+ This idea is used in most computer programs that implement maximum likelihood. In the Bayesian framework, the same idea was proposed in [Turek et al. (2016)](https://doi.org/10.1007/s10651-016-0353-z).

+ Cannot be done in Jags. Can be done in nimble thanks to nimble functions!

+ The run is *much* faster. Also allows fitting models to big datasets. More details in dedicated Worksheet.

---
background-color: #234f66
## <span style="color:white">No live demo, but there is a worksheet.</span>

.center[
![](img/b5b086f9cc403008ba7be5dd508cfed2.gif)
]

---
## Future directions for NIMBLE

+ NIMBLE is under active development.  Contributors are welcome, including those who want to get involved but don't know where.

+ Faster building of models and algorithms.  Ability to save and re-load compiled work.

+ Automatic differentiation of model calculations, enabling Hamiltonian Monte Carlo, other sampling strategies, and Laplace approximation.

+ Tools for building packages that use NIMBLE "under the hood".

---
## Further reading

+ Turek, D., de Valpine, P. & Paciorek, C.J. [Efficient Markov chain Monte Carlo sampling for hierarchical hidden Markov models](https://doi.org/10.1007/s10651-016-0353-z) *Environ Ecol Stat* 23: 549–564 (2016).

+ Ponisio, L.C., de Valpine, P., Michaud, N., and Turek, D. [One size does not fit all: Customizing MCMC methods for hierarchical models using NIMBLE](https://doi.org/10.1002/ece3.6053) *Ecol Evol.* 10: 2385–2416 (2020).

+ Nimble workshop to come 26-28 May, check out [here](https://r-nimble.org/nimble-virtual-short-course-may-26-28).

+ Nimble workshop material online available [here](https://github.com/nimble-training).

+ Nimble [manual](https://r-nimble.org/html_manual/cha-welcome-nimble.html) and [cheatsheet](https://r-nimble.org/cheatsheets/NimbleCheatSheet.pdf).