Default for :info ?

[MaximeBouton]

When testing RLInterface on POMDPs#version-0.8 I am getting the following error:
KeyError: key :info not found
which results from a call to:

s, r, info = gen(DDNOut(:sp, :r, :info), env.problem, env.state, a, env.rng)

I realized there is no default for :info anymore: https://github.com/JuliaPOMDP/POMDPModelTools.jl/blob/master/src/info.jl#L84

• Do we want a default that outputs nothing ?
• If yes, what is the proper way to implement it?
  e.g. POMDPs.gen(out::DDNOut{(:sp,:o,:r,:info)}, m, s, a, rng) = (gen(out, m, s, a, rng)..., nothing) , but with a smarter way to handle all the combination for out
• If not, what is the proper way to check if :info is available?

[zsunberg]

oh, hmm... I didn't think about this. I don't think it's a good idea to provide an implementation for gen. I guess the best way would be, in RLInterface

if :info in nodenames(DDNStructure(m))
    sp, r, i = gen(DDNOut(:sp, :r, :info), m, s, a, rng)
else
    sp, r = gen(DDNOut(:sp, :r), m, s, a, rng)
    i = nothing
end

I think the compiler should be able to remove the dead branch.

This is a little ugly, so maybe we should make something nicer eventually, like a DDNAll(m) that returns a NamedTuple, maybe even with defaults like DDNAll(m, info=nothing)

[MaximeBouton]

That is what I was looking for thanks. I don't think the compiler is able to do that since now it is breaking the type stability tests. I will try to find a better way.

[MaximeBouton]

I found a way to keep the type stability but it seems a bit hacky.
I added a parameter Info to MDPEnvironment that I set to true or false at construction if info is available.
See JuliaPOMDP/RLInterface.jl@f4668fc

[zsunberg]

Maybe in POMDPModelTools there should be a wrapper that adds a node if it doesn't exist, i.e.

wrapper = assure_hasnodes(pomdp, (:info=>nothing, :c=>0.0))

If pomdp already has :info and :c nodes, then it just passes it through; if it does not it creates a new POMDP that behaves like the old one, except it has nodes :info and :c with defaults nothing and 0.0 respectively.