Beyond `spades`

So far, we have run models with the spades function.

spades() only has a few arguments:

- debug
- .plotInitialTime
- .saveInitialTime
- progress
- cache

What if we want to do more stuff?

Replication

What if we have a stochastic model and need many independent (Monte Carlo) runs?

The experiment function does this.

?experiment # for details and examples

Generally, however, this creates problems because of the computational time required …

Replication

The replicates argument

It is as easy as replicates = 2 to do two replicates:

?experiment # for details

# See Example 5
#  ... run mySim from top of help file
sims <- experiment(mySim, replicates = 2)
attr(sims, "experiment")$expDesign # shows 2 reps of same expt

What does it return?

What is the attr?

The curse of replication

Replication is necessary but it multiplies the time it takes to run any model.

Fortunately, replication can use parallel processing:

This is the simplest, yet most effective, parallel processing;
No communication between individual simulations, until the end;
So they can be spread across threads on a single computer or among computers;
Requires a cluser (cl) argument.

Parallel

SpaDES functions that are parallel aware:
- experiment, POM, splitRaster
In each case, there are two ways:

## option 1:
raster::beginCluster() # make a cluster silently in background

## option 2:
cl <- parallel::makeCluster() # need to explicitly pass the
                              # cluster object into functions
                              # using cl argument

Try it!

`experiment`: Varying parameter values

?experiment
## see example 1
experimentParams <- list(fireSpread = 
                           list(spreadprob = list(0.2, 0.23),
                                nFires = list(20, 10)),
                         caribouMovement = 
                           list(N = list(100, 1000)))
sims <- experiment(mySim, params = experimentParams)
attr(sims, "experiment")

`experiment`: Saving files

Often, a spades call will generate output files. Where do they go when using experiment?

?experiment
## see example 4
sims <- experiment(mySim, 
                   params = experimentParams, 
                   dirPrefix = c("expt", "simNum"))
attr(sims, "experiment")$expVals # shows 8 alternative
               #experiment levels, 24 unique parameter values

dir(outputPath(mySim))

Working through other examples of `experiment`

Exercise:

Pick a few examples and try to understand them. (NOTE: they get more complicated as you get towards the bottom.)

?experiment

Running parallel simulations

Clearly, using experiment can take a lot of resources, but there are a few options:

Use your own computer, a local cluster, a server, cloud and more;
Access Compute Canada for free (though sometimes the queues are long):
- Blog post on using Compute Canada
- There is a different example here: ssh among Linux machines
  - Shows how to use a Linux cluster connected via SSH;
  - Can be very effective because is interactive (unlike normal compute clusters).

Pattern Oriented Modeling

Estimating unknown parameters

Ecological models use equations and functions
These have parameters
Parameters can be estimated from:
- directly from data (e.g., regression, machine learning)
- indirectly using (e.g., pattern oriented modeling) (e.g., Grimm et al. Science 2005, Marucco & McIntire J Appl Ecol 2010)

Directly from data

(Just a brief discussion of this here)
Remember, we are using R!
So you can use any statistical method that R has
Mixed models, non-linear models, machine learning, bayesian

nTrees <- 50
diamCM <- rlnorm(nTrees, meanlog = 2, sdlog = 0.7)
height <- rnorm(nTrees, sqrt(diamCM)*2 + rnorm(nTrees))
glmObj <- glm(height ~ diamCM)
plot(diamCM, height)
summary(glmObj)

Predict method

These methods often have predict methods associated with them
So, instead of using “fixed” parameters in simulation models
We can use these methods which give more accurate prediction:
- given uncertainty,
- random effects,
- covariance matrix

# predict with new data
predVals <- predict(glmObj, newdata = data.frame(
  diamCM = rlnorm(nTrees, meanlog = 2, sdlog = 0.7)
))

Using this in simulation models

In R, everything is an object
The glmObj and predictVals are objects
These have many features that can be used in simulations
- plot methods
- standard errors, covariance matrices
- random effects terms
For the first 25 years of simulation modeling in ecology, parameters were copied into simulation models and hard coded
This made it difficult to update the simulation model if there are new data and the regression needs to be re-fit
Using SpaDES we can now fit the data and use the parameters all inside the simulation

Indirectly from data