• FIx R-library extract files to deal with multi-line comments and !include statements.

Jeremy has R code that solves the issue with agents scalars so that tag recapture probabilities for generating unbiased observations.

Is your feature request related to a problem? Please describe.
I want the ABM to run an iteration simulating data, that is compatible with a particular generalised stock assessment SS, CASAL, MULTIFAN-CL.

Describe the solution you'd like
The things that will need to be done

• Create new reports that spit out simulated data that are compatible with a generalized package (SS, MULTFAN-CL)
• Create Harvest control rules, that take output from a generalized package output to create a new catch, that then gets run the next year to simulate data.
• create calls from within IBM to call estimation of other programs SS and Casal2

Describe alternatives you've considered
Need to demonstrate that this IBM is computationally efficient going forward, and maybe get some people on board to help flush out reports for generalized packages.

Additional context

have hit very poor performance recently, so that has forced me to check handling agents as vectors. Basically less memory mangement, we put a bool on an agent and check whether it is a valid agent and we can overwrite invalid agents. This will allow for fast random element lookup which is critical for mortality processes.

Is your feature request related to a problem? Please describe.
There can be serious computational savings to be made if we cache exp() calls on the agent class. Currently each growth call which n_years * n_time_steps*n_agents we more often than not call exp() on M and growth parameterss. For a model that doesn't time-vary this could be reduced to just n_agents.

Add threading on all processes where each cell can be threaded independently, only in movement there will have to be a bit more thought when splicing agents to the same destination cell, but a mutex with condition variable should be adequate.

Some scrappy code exists to add Sex, but this needs to be tidied up and tested

I have found out that the MSE functionality of the ABM relies on Rinside and Rcpp. When you compile the code with these dependencies they use local paths for dependent libraries. I either need to add these dependencies .dll or .so with the exe or, I am thinking for now just create a branch so for users that want to use this. They will need to compile the code in order to use it.

Notify subscribers creates a link between objects and allows an efficient way to update dependencies that are cached.
this will allow us to try the threading. With the knowledge that the right cache will be rebuild e.g time-varying selectivity or M.

This is just a space where people can add ideas (a kind of wish list), if the idea gains traction we can migrate to its own issue as not to clutter.

In Jeremy's and Nokome's ABM they release tags before harvesting, thus making it independent to harvesting. I think I will make this a user decision. Users can release tags before mortality events or during mortality events.

Currently the inputs for the other ABM are tags released and scanned proportions.

• tag release by area
• scanned for each year which is a proportion of total catch

Currently we do not distinguish any life history trait in the partition structure. So any specific process that effects a single life history 'could' (needs to be tested) be slow. For this example, say we have our full fishery model like in the example SpatialModel and we want to add a larval component. Say incorporate larval drift from currents for a 6 week period after recruitment.
Options to test

• leave partition structure as is and just iterate over everything and only apply process to the age class of interest.
• Create a specific larval partition and start building up the concepts of "categories" in the program, could be an overhead when moving individuals back into the main partition.

Add a selectivity to the movement processes, this can be done by checking if an agent is available to move before we do a multinomial draw on where it moves.

This will allow for age or length based movement.

currently I create a private vector of random numbers that each cell/thread builds and accesss, and so we only have to control for a race condition when populating the vector. Why do this is because the random number generator is a shared resource so we get segment fault for simultaneous access.

Scott has suggested an alternative which follows the code

Before get stuck in, iterate over world and calculate number of agents in world

Populate the single Global vector (rand_vec) with random numbers
cell_offsets= 0
for (i in rows)
for (j in cols)
random number = rand_vec[cell_offsets + agent_iter]

this way we only allocate memory once, and because of the business rules we will never be accessing the same element.

A concept for a new layer specifically related to the recruitment event. In the recruitment process users must define a layer of proportions where to see recruits. If we want a full life cycle model then I think age 0 individuals should be seeded proportional to the SSB over the spatial domain.

• create a new numeric layer that will convert population biomass or abundance into proportion layer over a user specific set of cells.

This means that recruits will be spatially proportional to the SSB at the point of recruitment. Which will also allow the recruitment spatial pattern to change over time with the SSB.

I have discovered that as soon as we go down the threading route we can never reproduce exact model runs (unless we run on a single thread) because we have no control over the order of thread execution and so random number chain will be executed in a different order each time. More of a note than anything, I am happy with this, it just means that in order to validate functionality, either run comparisons on single thread with a fixed seed or set agents to a massive number so that it would be obvious that the model is behaving differently to how you expect.

Add a subcommand to this process saying apply_selectivity. Many models assume constant selectivity, for this type of model2, we can ask agents to cache the exp(M) call. The process will need to find out how many times it is applied in the annual cycle as this will define the 'selectivity'. It might also want to cache for each time-step.

Add a flag on the model if doing a simulation whether initialization phase needs to be re-run.

Unless you are doing a -i with values such as B0, M or growth changing you shouldn't need to re-run the initialization phase. This would save quite a bit of time. Check if we are doing simulation, cache the world and then iterate over the model years for each simulation.