Merge pull request #7 from JuliaPOMDP/unweighted

added unweighted

src/ParticleFilters.jl

@@ -19,7 +19,8 @@ export
     SimpleParticleFilter,
     ImportanceResampler,
     LowVarianceResampler,
-    SIRParticleFilter
+    SIRParticleFilter,
+    UnweightedParticleFilter
 
 export
     resample,
@@ -49,7 +50,7 @@ abstract type AbstractParticleBelief{T} end
 # DEPRECATED: remove in future release
 Base.eltype{T}(::Type{AbstractParticleBelief{T}}) = T
 
-sampletype(::Type{AbstractParticleBelief{T}}) where T = T
+sampletype(::Type{B}) where B<:AbstractParticleBelief{T} where T = T
 
 ### Belief types ###
 
@@ -215,6 +216,7 @@ function SIRParticleFilter(model, n::Int; rng::AbstractRNG=Base.GLOBAL_RNG)
     return SimpleParticleFilter(model, LowVarianceResampler(n), rng)
 end
 
+include("unweighted.jl")
 include("beliefs.jl")
 include("updater.jl")
 include("resamplers.jl")

src/beliefs.jl

@@ -1,4 +1,3 @@
-
 n_particles(b::ParticleCollection) = length(b.particles)
 particles(p::ParticleCollection) = p.particles
 weighted_particles(p::ParticleCollection) = (p=>1.0/length(p.particles) for p in p.particles)

src/unweighted.jl

@@ -0,0 +1,44 @@
+"""
+UnweightedParticleFilter
+
+A particle filter that does not use any reweighting, but only keeps particles if the observation matches the true observation exactly. This does not require obs_weight, but it will not work well in real-world situations.
+"""
+struct UnweightedParticleFilter{M, RNG<:AbstractRNG} <: Updater
+    model::M
+    n::Int
+    rng::RNG
+end
+
+function UnweightedParticleFilter(model, n::Integer; rng=Base.GLOBAL_RNG)
+    return UnweightedParticleFilter(model, n, rng)
+end
+
+function update(up::UnweightedParticleFilter, b::ParticleCollection, a, o)
+    new = sampletype(b)[]
+    i = 1
+    while i <= up.n
+        s = particle(b, mod1(i, n_particles(b)))
+        sp, o_gen = generate_so(up.model, s, a, up.rng)
+        if o_gen == o
+            push!(new, sp)
+        end
+        i += 1
+    end
+    if isempty(new)
+        warn("""
+             Particle Depletion!
+
+             The UnweightedParticleFilter generated no particles consistent with observation $o. Consider upgrading to a SIRParticleFilter or a SimpleParticleFilter or creating your own domain-specific updater.
+             """
+            )
+    end
+    return ParticleCollection(new)
+end
+
+function update(up::UnweightedParticleFilter, b, a, o)
+    return update(up, initialize_belief(up, b), a, o)
+end
+
+function initialize_belief(up::UnweightedParticleFilter, b)
+    return ParticleCollection(collect(rand(up.rng, b) for i in 1:up.n))
+end

test/runtests.jl

@@ -32,3 +32,10 @@ m = mode(b)
 m = mean(b)
 it = iterator(b)
 weighted_particles(b)
+
+rng = MersenneTwister(47)
+uf = UnweightedParticleFilter(p, 1000, rng)
+ps = initialize_belief(uf, initial_state_distribution(p))
+a = rand(rng, actions(p))
+sp, o = generate_so(p, rand(rng, ps), a, rng)
+bp = update(uf, ps, a, o)