Solver package update checklist

[zsunberg]

To update a solver to v0.9, the following actions need to be performed:

• In Project.toml, update [compat] for POMDPs to include 0.9
• Use @gen instead of gen(DDNOut, ...)
• Use rand(rng, initialstate(m)) instead of initialstate(m, rng)
• Use initialstate(m) instead of initialstate_distribution(m)
• Use rand(rng, initialobs(m, s)) instead of initialobs(m, s, rng)
• Check for deprecation warnings in tests
• If the solver specifies requirements, use the versions from POMDPLinter (probably by importing them explicitly)

---

• DiscreteValueIteration [@zsunberg]

• LocalApproximationValueIteration

• GlobalApproximationValueIteration [@zsunberg]

• MCTS [@zsunberg ]

• QMDP [@zsunberg]

• FIB [@zsunberg]

• BeliefGridVI [@MaximeBouton]

• SARSOP (POMDPXFiles)

• BasicPOMCP [@zsunberg] (waiting on ParticleFilters)

• ARDESPOT [@zsunberg]

• MCVI [@zsunberg]

• POMDPSolve [@zsunberg]

• IncrementalPruning [@zsunberg]

• POMCPOW [@zsunberg]

• AEMS

• ParticleFilters [@lassepe]

• RLInterface [@MaximeBouton]

• DeepQLearning [@MaximeBouton]

• TabularTDLearning [@MaximeBouton]

• POMDPTesting

[MaximeBouton]

What about the new POMDPLinter? Should we use this for dependencies?

[zsunberg]

Good call. I will add it.

[MaximeBouton]

just a clarification for the compat , the @gen change is not backward compatible right? so should compat be:
POMDPs = "0.9" in that case?

[zsunberg]

No, I put @gen in 0.8.4 actually

…

On Tue, Aug 4, 2020 at 9:34 PM Maxime ***@***.***> wrote: just a clarification for the compat , the @gen change is not backward compatible right? so should compat be: POMDPs = "0.9" in that case? — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#312 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ABALI24B2DSYBZJBEBH3RG3R7DOLZANCNFSM4PF6FE4A> .

[zsunberg]

I mean it's fine to just update to only POMDPs = "0.9" if that is easiest, but it is also a possibility to do POMDPs = "0.8.4, 0.9" with @gen

[zsunberg]

Ok, I added a checked list to keep track of the solver updates. If you're willing to help, please put your name by an item and claim it!

[lassepe]

What about ParticleFilters.jl? Or is this tracked somewhere else (since it's not a solver)?

[zsunberg]

We should track it here.

…

On Wed, Aug 5, 2020 at 11:08 AM Lasse Peters ***@***.***> wrote: What about ParticleFilters.jl? Or is this tracked somewhere else (since it's not a solver)? — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#312 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ABALI2YRZUJWM7BEA7RKE33R7GN2LANCNFSM4PF6FE4A> .

[zsunberg]

@MaximeBouton thanks for doing SARSOP!

[zsunberg]

@lassepe a lot of the remaining solvers depend on ParticleFilters. did you have a chance to start updating it yet?

[lassepe]

Sorry for the delay guys. I got around to fix most of the things today but I am getting an error in two of the @implemented tests for obs_weight. I marked them as @test_broken for now. I'm not sure what is going on there. Did we change anything about the way obs_weight works?

Everything else should be done.

I can have another look at it tonight and figure out what is going on. But if you are faster than me @zsunberg or know immediately how to fix the obs_weight issue, feel free to jump in!

[zsunberg]

Thanks - you can just leave it as a test broken for now, and I will take a look at it

[lassepe]

I got around to fix it. The issue was that as of JuliaPOMDP/POMDPModelTools.jl@980076d there is only the 5 argument version of obs_weight provided by default now. Also it seems as if @implemented obs_weight(::P, ...) returns true even if for P <: POMDP neither observation nor obs_weight is implemented (?).

I removed the tests for the other methods of the obs_weight weight function.
Also, I found a bug in the travis config for ParticleFilters.jl (JuliaPOMDP/ParticleFilters.jl#39) which caused it to only ever build the docs but never run the test. This has been around for a while and we may want to check whether this has happened with any other package.

[lassepe]

Should I already tag/release/register a new version for ParticleFilters.jl?

[zsunberg]

Yeah... implemented has been a mess since POMDPs 0.8. It was never really that clean to begin with. When in doubt, implemented should just return missing. Avoid relying on it outside of POMDPLinter.

Yes, please register a new version!

[lassepe]

Okay, all done!

[zsunberg]

Thanks @lassepe this is much appreciated!

[zsunberg]

OK, we're done with the solvers!