Merge pull request #26 from JuliaPOMDP/pomdps-08-compat

POMDPs v0.8 Compatibility

.travis.yml

@@ -5,6 +5,7 @@ os:
 
 julia:
   - 1.0
+  - 1
 
 notifications:
   email: false

Project.toml

@@ -1,7 +1,7 @@
 name = "POMDPModelTools"
 uuid = "08074719-1b2a-587c-a292-00f91cc44415"
 authors = ["JuliaPOMDP Contributors"]
-version = "0.1.7"
+version = "0.2.0"
 
 [deps]
 Distributions = "31c24e10-a181-5473-b8eb-7969acd0382f"
@@ -14,14 +14,16 @@ UnicodePlots = "b8865327-cd53-5732-bb35-84acbb429228"
 
 [compat]
 Distributions = ">= 0.17"
+POMDPs = "0.7.3, 0.8.0"
 julia = "1"
 
 [extras]
+BeliefUpdaters = "8bb6e9a1-7d73-552c-a44a-e5dc5634aac4"
 POMDPModels = "355abbd5-f08e-5560-ac9e-8b5f2592a0ca"
 POMDPPolicies = "182e52fb-cfd0-5e46-8c26-fd0667c990f4"
 POMDPSimulators = "e0d0a172-29c6-5d4e-96d0-f262df5d01fd"
 Pkg = "44cfe95a-1eb2-52ea-b672-e2afdf69b78f"
 Test = "8dfed614-e22c-5e08-85e1-65c5234f0b40"
 
 [targets]
-test = ["Test", "POMDPModels", "POMDPSimulators", "POMDPPolicies", "Pkg"]
+test = ["Test", "POMDPModels", "POMDPSimulators", "POMDPPolicies", "BeliefUpdaters", "Pkg"]

docs/make.jl

@@ -1,17 +1,13 @@
+push!(LOAD_PATH, "../src/")
+
 using Documenter, POMDPModelTools
 
 makedocs(
     modules = [POMDPModelTools],
-    format = :html,
+    format = Documenter.HTML(),
     sitename = "POMDPModelTools.jl"
 )
 
 deploydocs(
     repo = "github.com/JuliaPOMDP/POMDPModelTools.jl.git",
-    julia = "1.0",
-    osname = "linux",
-    target = "build",
-    deps = nothing,
-    make = nothing
 )
-

docs/src/index.md

@@ -1,6 +1,6 @@
 # About
 
-POMDPModelTools is a collection of interface extensions and tools to make writing models and solvers for [POMDPs.jl](github.com/JuliaPOMDP/POMDPs.jl) easier.
+POMDPModelTools is a collection of interface extensions and tools to make writing models and solvers for [POMDPs.jl](https://github.com/JuliaPOMDP/POMDPs.jl) easier.
 
 ```@contents
 ```

docs/src/interface_extensions.md

@@ -35,9 +35,15 @@ ordered_observations
 
 It is often the case that useful information besides the belief, state, action, etc is generated by a function in POMDPs.jl. This information can be useful for debugging or understanding the behavior of a solver, updater, or problem. The info interface provides a standard way for problems, policies, solvers or updaters to output this information. The recording simulators from [POMDPSimulators.jl](https://github.com/JuliaPOMDP/POMDPSimulators.jl) automatically record this information.
 
+To specify info for a problem (in POMDPs v0.8 and above), one should modify the problem's DDN with the `add_infonode` function, then return the info in `gen`. There is an example of this pattern in the docstring below:
+
+```@docs
+add_infonode
+```
+
+To specify info from policies, solvers, or updaters, implement the following functions:
+
 ```@docs
-generate_sri
-generate_sori
 action_info
 solve_info
 update_info

src/POMDPModelTools.jl

@@ -6,9 +6,9 @@ using LinearAlgebra
 using SparseArrays
 using UnicodePlots
 
-import POMDPs: actions, n_actions, actionindex
-import POMDPs: states, n_states, stateindex
-import POMDPs: observations, n_observations, obsindex
+import POMDPs: actions, actionindex
+import POMDPs: states, stateindex
+import POMDPs: observations, obsindex
 import POMDPs: sampletype, generate_sr, initialstate, isterminal, discount
 import POMDPs: implemented
 import Distributions: pdf, mode, mean, support
@@ -22,11 +22,12 @@ include("visualization.jl")
 
 # info interface
 export
-    generate_sri,
-    generate_sori,
+    add_infonode,
     action_info,
     solve_info,
-    update_info
+    update_info,
+    generate_sri,
+    generate_sori
 include("info.jl")
 
 export

src/distributions/bool.jl

@@ -23,11 +23,6 @@ function Base.iterate(d::BoolDistribution, state::Bool)
 end    
 
 support(d::BoolDistribution) = [true, false]
-
-==(d1::BoolDistribution, d2::BoolDistribution) = d1.p == d2.p
-
-Base.hash(d::BoolDistribution) = hash(d.p)
-
 Base.length(d::BoolDistribution) = 2
 
 Base.show(io::IO, m::MIME"text/plain", d::BoolDistribution) = showdistribution(io, m, d, title="BoolDistribution")

src/distributions/deterministic.jl

@@ -13,6 +13,7 @@ rand(rng::AbstractRNG, d::Deterministic) = d.val
 rand(d::Deterministic) = d.val
 support(d::Deterministic) = (d.val,)
 sampletype(::Type{Deterministic{T}}) where T = T
+Random.gentype(::Type{Deterministic{T}}) where T = T
 pdf(d::Deterministic, x) = convert(Float64, x == d.val)
 mode(d::Deterministic) = d.val
 mean(d::Deterministic{N}) where N<:Number = d.val / 1 # / 1 is to make this return a similar type to Statistics.mean

src/distributions/sparse_cat.jl

@@ -90,6 +90,7 @@ end
 Base.length(d::SparseCat) = min(length(d.vals), length(d.probs))
 Base.eltype(D::Type{SparseCat{V,P}}) where {V, P} = Pair{eltype(V), eltype(P)}
 sampletype(D::Type{SparseCat{V,P}}) where {V, P} = eltype(V)
+Random.gentype(D::Type{SparseCat{V,P}}) where {V, P} = eltype(V)
 
 function mean(d::SparseCat)
     vsum = zero(eltype(d.vals))

src/distributions/uniform.jl

@@ -24,6 +24,7 @@ end
 
 support(d::Uniform) = d.set
 sampletype(::Type{Uniform{T}}) where T = eltype(T)
+Random.gentype(::Type{Uniform{T}}) where T = eltype(T)
 
 function pdf(d::Uniform, s)
     if s in d.set
@@ -49,6 +50,7 @@ end
 pdf(d::UnsafeUniform, s) = 1.0/length(d.collection)
 support(d::UnsafeUniform) = d.collection
 sampletype(::Type{UnsafeUniform{T}}) where T = eltype(T)
+Random.gentype(::Type{UnsafeUniform{T}}) where T = eltype(T)
 
 # Common Implementations
 

src/fully_observable_pomdp.jl

@@ -3,24 +3,33 @@
 
 Turn `MDP` `mdp` into a `POMDP` where the observations are the states of the MDP.
 """
-struct FullyObservablePOMDP{S, A} <: POMDP{S,A,S}
-    mdp::MDP{S, A}
+struct FullyObservablePOMDP{M,S,A} <: POMDP{S,A,S}
+    mdp::M
 end
 
+function FullyObservablePOMDP(m::MDP)
+    return FullyObservablePOMDP{typeof(m), statetype(m), actiontype(m)}(m)
+end
+
+mdptype(::Type{FullyObservablePOMDP{M,S,A}}) where {M,S,A} = M
+
+function POMDPs.DDNStructure(::Type{M}) where M <: FullyObservablePOMDP
+    MM = mdptype(M)
+    add_obsnode(DDNStructure(MM))
+end
+
+add_obsnode(ddn) = add_node(ddn, :o, FunctionDDNNode((m,sp)->sp), (:sp,)) # for ::DDNStructure, but this is not declared yet POMDPs in v0.7.3
+
 POMDPs.observations(pomdp::FullyObservablePOMDP) = states(pomdp.mdp)
-POMDPs.n_observations(pomdp::FullyObservablePOMDP) = n_states(pomdp.mdp)
 POMDPs.obsindex(pomdp::FullyObservablePOMDP{S, A}, o::S) where {S, A} = stateindex(pomdp.mdp, o)
 
 POMDPs.convert_o(T::Type{V}, o, pomdp::FullyObservablePOMDP) where {V<:AbstractArray} = convert_s(T, s, pomdp.mdp)
 POMDPs.convert_o(T::Type{S}, vec::V, pomdp::FullyObservablePOMDP) where {S,V<:AbstractArray} = convert_s(T, vec, pomdp.mdp)
 
+POMDPs.gen(::DDNNode{:o}, m::FullyObservablePOMDP, sp, rng) = sp
 
-function POMDPs.generate_o(pomdp::FullyObservablePOMDP, s, a, rng::AbstractRNG)
-    return s
-end
-
-function POMDPs.observation(pomdp::FullyObservablePOMDP, s, a)
-    return Deterministic(s)
+function POMDPs.observation(pomdp::FullyObservablePOMDP, a, sp)
+    return Deterministic(sp)
 end
 
 function POMDPs.observation(pomdp::FullyObservablePOMDP, s, a, sp)
@@ -31,19 +40,30 @@ end
 
 POMDPs.states(pomdp::FullyObservablePOMDP) = states(pomdp.mdp)
 POMDPs.actions(pomdp::FullyObservablePOMDP) = actions(pomdp.mdp)
-POMDPs.transition(pomdp::FullyObservablePOMDP{S,A}, s::S, a::A) where {S,A} = transition(pomdp.mdp, s, a)
+POMDPs.transition(pomdp::FullyObservablePOMDP, s, a) = transition(pomdp.mdp, s, a)
 POMDPs.initialstate_distribution(pomdp::FullyObservablePOMDP) = initialstate_distribution(pomdp.mdp)
 POMDPs.initialstate(pomdp::FullyObservablePOMDP, rng::AbstractRNG) = initialstate(pomdp.mdp, rng)
-POMDPs.generate_s(pomdp::FullyObservablePOMDP, s, a, rng::AbstractRNG) = generate_s(pomdp.mdp, s, a, rng)
-POMDPs.generate_sr(pomdp::FullyObservablePOMDP, s, a, rng::AbstractRNG) = generate_sr(pomdp.mdp, s, a, rng)
-POMDPs.reward(pomdp::FullyObservablePOMDP{S, A}, s::S, a::A) where {S,A} = reward(pomdp.mdp, s, a)
 POMDPs.isterminal(pomdp::FullyObservablePOMDP, s) = isterminal(pomdp.mdp, s)
 POMDPs.discount(pomdp::FullyObservablePOMDP) = discount(pomdp.mdp)
-POMDPs.n_states(pomdp::FullyObservablePOMDP) = n_states(pomdp.mdp)
-POMDPs.n_actions(pomdp::FullyObservablePOMDP) = n_actions(pomdp.mdp)
-POMDPs.stateindex(pomdp::FullyObservablePOMDP{S,A}, s::S) where {S,A} = stateindex(pomdp.mdp, s)
-POMDPs.actionindex(pomdp::FullyObservablePOMDP{S, A}, a::A) where {S,A} = actionindex(pomdp.mdp, a)
+POMDPs.stateindex(pomdp::FullyObservablePOMDP, s) = stateindex(pomdp.mdp, s)
+POMDPs.actionindex(pomdp::FullyObservablePOMDP, a) = actionindex(pomdp.mdp, a)
 POMDPs.convert_s(T::Type{V}, s, pomdp::FullyObservablePOMDP) where V<:AbstractArray = convert_s(T, s, pomdp.mdp)
 POMDPs.convert_s(T::Type{S}, vec::V, pomdp::FullyObservablePOMDP) where {S,V<:AbstractArray} = convert_s(T, vec, pomdp.mdp)
 POMDPs.convert_a(T::Type{V}, a, pomdp::FullyObservablePOMDP) where V<:AbstractArray = convert_a(T, a, pomdp.mdp)
 POMDPs.convert_a(T::Type{A}, vec::V, pomdp::FullyObservablePOMDP) where {A,V<:AbstractArray} = convert_a(T, vec, pomdp.mdp)
+
+POMDPs.gen(d::DDNNode, m::FullyObservablePOMDP, args...) = gen(d, m.mdp, args...)
+POMDPs.gen(m::FullyObservablePOMDP, s, a, rng) = gen(m.mdp, s, a, rng)
+POMDPs.reward(pomdp::FullyObservablePOMDP, s, a) = reward(pomdp.mdp, s, a)
+
+# deprecated in POMDPs v0.8
+add_obsnode(ddn::POMDPs.DDNStructureV7{(:s,:a,:sp,:r)}) = POMDPs.DDNStructureV7{(:s,:a,:sp,:o,:r)}()
+add_obsnode(ddn::POMDPs.DDNStructureV7) = error("FullyObservablePOMDP only supports MDPs with the standard DDN Structure (DDNStructureV7{(:s,:a,:sp,:r)}) with POMDPs v0.7.")
+
+POMDPs.generate_s(pomdp::FullyObservablePOMDP, s, a, rng::AbstractRNG) = generate_s(pomdp.mdp, s, a, rng)
+POMDPs.generate_sr(pomdp::FullyObservablePOMDP, s, a, rng::AbstractRNG) = generate_sr(pomdp.mdp, s, a, rng)
+POMDPs.n_actions(pomdp::FullyObservablePOMDP) = n_actions(pomdp.mdp)
+POMDPs.n_states(pomdp::FullyObservablePOMDP) = n_states(pomdp.mdp)
+function POMDPs.generate_o(pomdp::FullyObservablePOMDP, s, rng::AbstractRNG)
+    return s
+end

src/generative_belief_mdp.jl

@@ -14,15 +14,20 @@ function GenerativeBeliefMDP(pomdp::P, up::U) where {P<:POMDP, U<:Updater}
     GenerativeBeliefMDP{P, U, typeof(b0), actiontype(pomdp)}(pomdp, up)
 end
 
-function generate_sr(bmdp::GenerativeBeliefMDP, b, a, rng::AbstractRNG)
+function POMDPs.gen(bmdp::GenerativeBeliefMDP, b, a, rng::AbstractRNG)
     s = rand(rng, b)
     if isterminal(bmdp.pomdp, s)
         bp = gbmdp_handle_terminal(bmdp.pomdp, bmdp.updater, b, s, a, rng::AbstractRNG)::typeof(b)
         return bp, 0.0
     end
-    sp, o, r = generate_sor(bmdp.pomdp, s, a, rng) # maybe this should have been generate_or?
+    sp, o, r = gen(DDNOut(:sp,:o,:r), bmdp.pomdp, s, a, rng) # maybe this should have been generate_or?
     bp = update(bmdp.updater, b, a, o)
-    return bp, r
+    return (sp=bp, r=r)
+end
+
+function generate_sr(bmdp::GenerativeBeliefMDP, b, a, rng::AbstractRNG)
+    x = gen(bmdp, b, a, rng)
+    return x.sp, x.r
 end
 
 function initialstate(bmdp::GenerativeBeliefMDP, rng::AbstractRNG)

src/info.jl

@@ -1,24 +1,6 @@
 # functions for passing out info from simulations, similar to the info return from openai gym
 # maintained by @zsunberg
 
-"""
-Return a tuple containing the next state and reward and information (usually a `NamedTuple`, `Dict` or `nothing`) from that step.
-
-By default, returns `nothing` as info.
-"""
-function generate_sri(p::MDP, s, a, rng::AbstractRNG)
-    return generate_sr(p, s, a, rng)..., nothing
-end
-
-"""
-Return a tuple containing the next state, observation, and reward and information (usually a `NamedTuple`, `Dict` or `nothing`) from that step.
-
-By default, returns `nothing` as info. 
-"""
-function generate_sori(p::POMDP, s, a, rng::AbstractRNG)
-    return generate_sor(p, s, a, rng)..., nothing
-end
-
 """
     a, ai = action_info(policy, x)
 
@@ -51,3 +33,78 @@ By default, returns `nothing` as info.
 function update_info(up::Updater, b, a, o)
     return update(up, b, a, o), nothing
 end
+
+# once POMDPs v0.8 is released, this should be a jldoctest
+"""
+    add_infonode(ddn::DDNStructure)
+
+Create a new DDNStructure object with a new node labeled :info for returning miscellaneous informationabout a simulation step.
+
+Typically, the object in info is associative (i.e. a `Dict` or `NamedTuple`) with keys corresponding to different pieces of information.
+
+# Example (using POMDPs v0.8)
+
+```julia
+using POMDPs, POMDPModelTools, POMDPPolicies, POMDPSimulators, Random
+
+struct MyMDP <: MDP{Int, Int} end
+
+# add the info node to the DDN
+POMDPs.DDNStructure(::Type{MyMDP}) = mdp_ddn() |> add_infonode
+
+# the dynamics involve two random numbers - here we record the values for each in info
+function POMDPs.gen(m::MyMDP, s, a, rng)
+    r1 = rand(rng)
+    r2 = randn(rng)
+    return (sp=s+a+r1+r2, r=s^2, info=(r1=r1, r2=r2))
+end
+
+m = MyMDP()
+@show nodenames(DDNStructure(m))
+p = FunctionPolicy(s->1)
+for (s,info) in stepthrough(m, p, 1, "s,info", max_steps=5, rng=MersenneTwister(2))
+    @show s
+    @show info 
+end
+```
+"""
+function add_infonode(ddn) # for DDNStructure, but it is not declared in v0.7.3, so there is not annotation
+    add_node(ddn, :info, ConstantDDNNode(nothing), nodenames(ddn))
+end
+
+function add_infonode(ddn::POMDPs.DDNStructureV7{nodenames}) where nodenames
+    return POMDPs.DDNStructureV7{(nodenames..., :info)}()
+end
+
+###############################################################
+# Note all generate functions will be deprecated in POMDPs v0.8
+###############################################################
+
+
+if DDNStructure(MDP) isa POMDPs.DDNStructureV7
+    """
+    Return a tuple containing the next state and reward and information (usually a `NamedTuple`, `Dict` or `nothing`) from that step.
+
+    By default, returns `nothing` as info.
+    """
+    function generate_sri(p::MDP, s, a, rng::AbstractRNG)
+        return generate_sr(p, s, a, rng)..., nothing
+    end
+
+    """
+    Return a tuple containing the next state, observation, and reward and information (usually a `NamedTuple`, `Dict` or `nothing`) from that step.
+
+    By default, returns `nothing` as info. 
+    """
+    function generate_sori(p::POMDP, s, a, rng::AbstractRNG)
+        return generate_sor(p, s, a, rng)..., nothing
+    end
+
+    POMDPs.gen(::DDNOut{(:sp,:o,:r,:i)}, m, s, a, rng) = generate_sori(m, s, a, rng)
+    POMDPs.gen(::DDNOut{(:sp,:o,:r,:info)}, m, s, a, rng) = generate_sori(m, s, a, rng)
+    POMDPs.gen(::DDNOut{(:sp,:r,:i)}, m, s, a, rng) = generate_sri(m, s, a, rng)
+    POMDPs.gen(::DDNOut{(:sp,:r,:info)}, m, s, a, rng) = generate_sri(m, s, a, rng)
+else
+    @deprecate generate_sri(args...) gen(DDNOut(:sp,:r,:info), args...)
+    @deprecate generate_sori(args...) gen(DDNOut(:sp,:o,:r,:info), args...)
+end