Is our instructions on not using MultiAgent models really true?

[Datseris]

I mean, sure, if the agent stepping function takes nanoseconds to microseconds to run, obviously type stability is a big hit to performance. However our existing benchmarkings are for a rather trivial agent stepping dynamics. If the agent step function itself takes significant amount of time, the dispatch cost becomes meaningless.

This is rather important because the multi-agent approach leads to dramatically simpler code. I'm just fearing that we might be pushing users too much in the direction "avoid multi agent models" with the current documentation status.

[Libbum]

A fair point. We do have multiagent benchmarks that show a slowdown, but as you say—they are for trivial cases.

Multi-minute or hour long processes are something we're not testing at present. Let's generate a few things to that effect.

[AayushSabharwal]

Maybe we can just use predator-prey? It already has a multiagent version, and sensible stepping functions

[veddox]

This would be interesting to know! Does anybody have any new data on this?

My model will have dozens of species, and at the moment I'm using a work-around with trait dicts assigned to a single animal agent type to avoid creating numerous types. If this is not such a big problem after all, I could probably write some much nicer code.

[Datseris]

You should make a minimal model and test it out. One with one type and a trait field and one with many types. Let us know!

[jmaspons]

What's the status of multiagent? I found a comment on deprecation in the code

Agents.jl/src/submodules/schedulers.jl

Line 214 in 14a36df

# TODO: ByKind scheduler should be removed when deprecation of @multiagent is complete

[Tortar]

Hi @jmaspons, I think this issue is outdated, @multiagent is now faster than multiple types in many cases and mostly on par in others. But it is true that the ByKind scheduler is outdated, it was there from a previous iteration of the multiagent infrastructure, it should be removed and the ByType one be updated to support multiagents, or alternatively, created a new scheduler e.g. ByVariant specifically for multiagents

[Datseris]

I am closing this issue because it refers to an outdated docmentation discussion. Our current docs are rather clear on the pros and cons of the various multi agent approaches and also promote @multiagent as much as possible. Regarding the ByKind being replacedwith ByVariant, this should be a new issue (or even better directly a new Pull Request given the simplicity of the change)