Automated Pattern Oriented Modeling

[eliotmcintire]

There are several ways to build a tool to created an automated fitting procedure.

Requirements for user:

1. Data that matches with some of the outputs from simulation
2. Parameters that can vary
3. Objective (cost) function

On the SpaDES side, we need either one of two mechanisms:

1. A module that when added, it does the optimization, or
2. A function in which the user is more involved in the optimization

Likely, it will use the experiment function internally, allowing for multithreading, though some of the optimization functions in R will do this internally. DEoptim function in DEoptim package does internal parallel processing.

Advantages of module version - it can be done by a non-Power-R-user, i.e., they don't have to know the scripting part of setting up the optimization. This would allow the optimization to be done internally within a spades call, thus, can be wrapped into a shiny GUI.

Advantages of function version - it can only be done by a Power R user because of the necessity of scripting and data munging/wrangling to get things in place.

Likely, it will be both, i.e., the module will wrap the function.

[eliotmcintire]

@ygc2l Any news on updated POM function?

[ygc2l]

Nope. I gave up on using it, handling a custom objective function is not what I think this function should be made for. Using POM() for a problem that require a custom objective function, i.e. something that is not a SpaDES module, reduce POM() as a wrapper and would somewhat complicate things unnecessarily for the end-user. But maybe you have another opinion on this ?

[eliotmcintire]

So what is your solution for your specific problem? Can you outline it
here? I believe that there is value for such a function because parameters
are known and objects are known and data are known (all known from module
metadata). So, rather than a user needing to know 4 separate things, they
only need one. And there are a small set of defaults that would be useful.

[ygc2l]

In my case, I only needed the spread function to perform the simulations not a whole SpaDES module. My solution was to 1) create a SpaDES module to do the optimization (not the simulation), 2) design the module parameters to be used inside the objective function, and 3) set data requirements as inputs for the module.

If I had to use POM() to do this, I would still have to supply the objective function, the parameters of DEoptim and the data. Thus, I don't see how I would gain something by using it.

Btw, I don't quite understand what your saying by "So, rather than a user needing to know 4 separate things, they only need one."

[achubaty]

this is mostly done. the only thing remaining is to implement the changes arising from the forthcoming update to DEoptim, but there is no solid timeline from them, so I am closing this issue and prepping for CRAN release of v1.3.0.