Created
August 18, 2016 11:22
-
-
Save pkofod/41ef3bceab07586e62a2cd01590bc56e to your computer and use it in GitHub Desktop.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
function LearnBase.learn!(solver::CrossEntropyMethod, env::AbstractEnvironment, doanim = false) | |
# !!! INIT: | |
# this is a mappable function of θ to reward | |
cem_episode = θ -> begin | |
π = cem_policy(env, θ) | |
R, T = episode!(env, π; maxiter = solver.options[:maxiter]) | |
R | |
end | |
result = optimize(cem_episode, solver.μ, CrossEntropy()) | |
# is solver.μ the paramter to be found? | |
solver.μ[:] = Optim.minimzer(result) | |
end | |
method_string(method::CrossEntropy) = "Nicely Formatted Method" | |
type CrossEntropyState{T} | |
# your variables if any | |
# default variables; need to document this, initial_x | |
end | |
initialize_state(method::Method, options, d, initial_x::Array) = initialize_state(method, options, d.f, initial_x) | |
function initialize_state(method::Method, options, f::Function, initial_x) | |
#something like | |
anim = doanim ? Animation() : nothing | |
n_elite = round(Int, solver.options[:cem_batch_size] * solver.options[:cem_elite_frac]) | |
last_μ = similar(solver.μ) | |
tr = OptimizationTrace{typeof(Method)}() | |
CrossEntropy(anim, n_elite, last_μ, tr, default_variables...) | |
end | |
update!(d, state::ParticleSwarmState, method::ParticleSwarm) = update!(d.f, state, method) | |
function update!{T}(f::Function, state::ParticleSwarmState{T}, method::ParticleSwarm) | |
last_μ = copy(solver.μ) | |
# sample thetas from a multivariate normal distribution | |
N = MultivariateNormal(solver.μ, solver.σ) | |
θs = [rand(N) for k=1:solver.options[:cem_batch_size]] | |
# compute rewards and pick out an elite set | |
Rs = map(cem_episode, θs) | |
elite_indices = sortperm(Rs, rev=true)[1:n_elite] | |
elite_θs = θs[elite_indices] | |
info("Iteration $t. mean(R): $(mean(Rs)) max(R): $(maximum(Rs))") | |
# update the policy from the elite set | |
for j=1:length(solver.μ) | |
θj = [θ[j] for θ in elite_θs] | |
solver.μ[j] = mean(θj) | |
solver.Z[j] = solver.noise_func(t) | |
solver.σ[j] = sqrt(var(θj) + solver.Z[j]) | |
end | |
@show solver.μ solver.σ solver.Z | |
state.R, state.T = episode!( | |
env, | |
cem_policy(env, solver.μ), | |
maxiter = solver.options[:maxiter], | |
stepfunc = myplot(t, hist_min, hist_mean, hist_max, anim) | |
) | |
false | |
end | |
function assess_convergence(state::CrossEntropy, options) | |
normdiff = norm(solver.μ - last_μ) | |
@show normdiff | |
if normdiff < state.options[:stopping_norm] | |
info("Converged after $(t*state.options[:cem_batch_size]) episodes.") | |
return true | |
end | |
false | |
end | |
function trace!(tr, state, iteration, method, options) | |
# save the three values in the extended trace using a dictionary | |
end | |
after_while!(d, state, method, options) | |
state.doanim && gif(state.anim) | |
end |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment