1
0
Fork 0
mirror of https://gitlab.rlp.net/mobitar/ReCo.jl.git synced 2024-09-17 18:51:17 +00:00

Added reward discount analysis

This commit is contained in:
Mo8it 2022-01-30 02:28:34 +01:00
parent a57bebbcb7
commit b5767f0104
4 changed files with 65 additions and 11 deletions

View file

@ -160,7 +160,7 @@ function plot_radial_distributions(;
max_g = maximum(maximum.(gs))
ax = Axis(
fig[1:2, 1:2];
fig[1, 1];
xticks=0:(2 * particle_radius):floor(Int64, max_lower_radius),
yticks=0:ceil(Int64, max_g),
xlabel=L"r / d",

View file

@ -0,0 +1,49 @@
using CairoMakie
using LaTeXStrings: @L_str
using ReCo: ReCo
includet("../src/Visualization/common_CairoMakie.jl")
function run_reward_discount_analysis()
γs = 0.0:0.2:1.0
n_γs = length(γs)
γ_rewards = Vector{Vector{Float64}}(undef, n_γs)
Threads.@threads for γ_ind in 1:n_γs
γ = γs[γ_ind]
env_helper = ReCo.run_rl(;
EnvType=ReCo.OriginEnv,
n_episodes=500,
episode_duration=8.0,
n_particles=200,
update_actions_at=0.08,
ϵ_stable=0.00001,
process_dir="reward_discount_analysis/$γ_ind",
reward_discount=γ,
)
rewards = env_helper.shared.hook.rewards
γ_rewards[γ_ind] = rewards
end
init_cairomakie!()
fig = gen_figure()
ax = Axis(fig[1, 1]; xlabel="Episode", ylabel="Reward")
rewards_plots = []
for (rewards_ind, rewards) in enumerate(γ_rewards)
rewards_plot = lines!(ax, 1:length(rewards), rewards)
push!(rewards_plots, rewards_plot)
end
Legend(fig[1, 2], rewards_plots, [L"\gamma = %$γ" for γ in γs])
set_gaps!(fig)
save_fig("reward_discount_analysis.pdf", fig)
return nothing
end

View file

@ -26,7 +26,9 @@ include("States.jl")
include("Hooks.jl")
include("Reward.jl")
function gen_agent(n_states::Int64, n_actions::Int64, ϵ_stable::Float64)
function gen_agent(
n_states::Int64, n_actions::Int64, ϵ_stable::Float64, reward_discount::Float64
)
# TODO: Optimize warmup and decay
warmup_steps = 500_000
decay_steps = 5_000_000
@ -36,7 +38,7 @@ function gen_agent(n_states::Int64, n_actions::Int64, ϵ_stable::Float64)
approximator=TabularQApproximator(;
n_state=n_states, n_action=n_actions, opt=Flux.InvDecay(1.0)
),
γ=0.95, # Reward discount
γ=reward_discount,
),
explorer=EpsilonGreedyExplorer(;
kind=:linear,
@ -67,6 +69,7 @@ function run_rl(;
skin_to_interaction_radius_ratio::Float64=ReCo.DEFAULT_SKIN_TO_INTERACTION_R_RATIO,
packing_ratio::Float64=0.15,
show_progress::Bool=true,
reward_discount::Float64=1.0,
) where {E<:Env}
@assert 0.0 <= elliptical_a_b_ratio <= 1.0
@assert n_episodes > 0
@ -89,7 +92,7 @@ function run_rl(;
env_args = (skin_radius=sim_consts.skin_radius, half_box_len=sim_consts.half_box_len)
env = EnvType(; args=env_args)
agent = gen_agent(env.shared.n_states, env.shared.n_actions, ϵ_stable)
agent = gen_agent(env.shared.n_states, env.shared.n_actions, ϵ_stable, reward_discount)
n_steps_before_actions_update = round(Int64, update_actions_at / sim_consts.δt)

View file

@ -1,7 +1,9 @@
using DataFrames
using DataFrames: DataFrames
using PrettyTables: pretty_table
function latex_table(dataframe, filename::String; path="exports/$filename")
function latex_table(
dataframe::DataFrames.DataFrame, filename::String; path::String="exports/$filename"
)
open(path, "w") do f
pretty_table(f, dataframe; backend=:latex, nosubheader=true, alignment=:c)
end
@ -9,8 +11,8 @@ function latex_table(dataframe, filename::String; path="exports/$filename")
return nothing
end
function latex_rl_table(env_helper, filename)
table = env_helper.shared.agent.policy.learner.approximator.table
function latex_rl_table(env_helper, filename::String)
table = copy(env_helper.shared.agent.policy.learner.approximator.table)
for col in 1:size(table)[2]
table[:, col] ./= sum(table[:, col])
@ -31,12 +33,12 @@ function latex_rl_table(env_helper, filename)
for i in action_spaces_labels[1]
for j in action_spaces_labels[2]
push!(actions, i * "; " * j)
push!(actions, i * " ; " * j)
end
end
df = DataFrame(table, states)
insertcols!(df, 1, :Actions => actions)
df = DataFrames.DataFrame(table, states)
DataFrames.insertcols!(df, 1, :Actions => actions)
latex_table(df, filename)