Parallel

The run_parallel function can be used to conveniently run simulations in parallel. Example:

using POMDPModels
using POMDPPolicies
using POMDPSimulators

pomdp = BabyPOMDP()
fwc = FeedWhenCrying()
rnd = solve(RandomSolver(MersenneTwister(7)), pomdp)

q = [] # vector of the simulations to be run
push!(q, Sim(pomdp, fwc, max_steps=32, rng=MersenneTwister(4), metadata=Dict(:policy=>"feed when crying")))
push!(q, Sim(pomdp, rnd, max_steps=32, rng=MersenneTwister(4), metadata=Dict(:policy=>"random")))

# this creates two simulations, one with the feed-when-crying policy and one with a random policy

data = run_parallel(q)

# by default, the dataframe output contains the reward and the contents of `metadata`
@show data
# data = 2×2 DataFrames.DataFrame
# │ Row │ policy             │ reward   │
# ├─────┼────────────────────┼──────────┤
# │ 1   │ "feed when crying" │ -4.5874  │
# │ 2   │ "random"           │ -27.4139 │

# to perform additional analysis on each of the simulations one can define a processing function with the `do` syntax:
data2 = run_parallel(q, progress=false) do sim, hist
println("finished a simulation - final state was $(last(state_hist(hist)))")
return [:steps=>n_steps(hist), :reward=>discounted_reward(hist)]
end

@show data2
# 2×3 DataFrames.DataFrame
# │ Row │ policy             │ reward   │ steps │
# ├─────┼────────────────────┼──────────┼───────┤
# │ 1   │ "feed when crying" │ -18.2874 │ 32.0  │
# │ 2   │ "random"           │ -17.7054 │ 32.0  │

run_parallel(queue::Vector{Sim})
run_parallel(f::Function, queue::Vector{Sim})

run_parallel(queue) do sim, hist
    return (n_steps=n_steps(hist), reward=discounted_reward(hist))
end