Transition function return values

[matthhong]

Hey Zach,

Referencing these lines:
https://github.com/JuliaPOMDP/QuickPOMDPs.jl/blob/master/src/discrete_explicit.jl#L162

Currently the inner for xp in ss loop evaluates the transition probabilities for each state-action combo, and populates a SparceCat distribution.

This creates a pretty big bottleneck, especially if the transition function itself is costly.

In your other models (e.g. https://github.com/JuliaPOMDP/POMDPModels.jl/blob/master/src/TigerPOMDPs.jl), T doesn't take a future state as a param. So they just return Distribution objects.

In discrete_explicit.jl, would it be possible to allow two different signatures for T (one taking s, a, and sp as inputs, and another taking only s and a), so it can skip the last inner loop when T returns a Distribution object?

If so, I could create a PR for it? Seems pretty straightforward.

[zsunberg]

Hey, thanks for thinking of this. There are actually two interfaces here - the QuickPOMDPs interface and the DiscreteExplicit interface.

In the QuickPOMDPs interface, you do write functions that return distributions. For example, https://github.com/JuliaPOMDP/QuickPOMDPs.jl/blob/master/examples/lightdark.jl

The DiscreteExplicit interface was really only meant for very simple problems, and even for those, as we found in class, it is fairly unweildy. I think maybe I should move it to another package to reduce confusion.