Added reward normalization

Project.toml

@@ -1,7 +1,7 @@
 name = "ReCo"
 uuid = "b25f7548-fcc9-4c91-bc24-841b54f4dd54"
 authors = ["MoBit <mo8it@protonmail.com>"]
-version = "0.3.0"
+version = "0.4.0"
 
 [deps]
 BenchmarkTools = "6e4b80f9-dd63-53aa-95a3-0cdb28fa8baf"
@@ -22,6 +22,7 @@ LinearAlgebra = "37e2e46d-f89d-539d-b4ee-838fcccc9c8e"
 LoopVectorization = "bdcacae8-1622-11e9-2a5c-532679323890"
 Luxor = "ae8d54c2-7ccd-5906-9d76-62fc9837b5bc"
 MathTeXEngine = "0a4f8689-d25c-4efe-a92b-7142dfc1aa53"
+PrettyTables = "08abe8d2-0d0c-5749-adfa-8a2ac140af0d"
 ProfileView = "c46f51b8-102a-5cf2-8d2c-8597cb0e0da7"
 ProgressMeter = "92933f4c-e287-5a05-a399-4b506db050ca"
 Random = "9a3f8284-a2c9-5f02-9a11-845980a1fd5c"

src/RL/EnvHelper.jl

@@ -6,6 +6,7 @@ struct EnvHelperSharedProps{H<:AbstractHook}
     hook::H
 
     n_steps_before_actions_update::Int64
+    n_actions_updates_per_episode::Int64
 
     elliptical_a_b_ratio::Float64
 
@@ -22,6 +23,7 @@ struct EnvHelperSharedProps{H<:AbstractHook}
         agent::Agent,
         hook::H,
         n_steps_before_actions_update::Int64,
+        n_actions_updates_per_episode::Int64,
         elliptical_a_b_ratio::Float64,
         n_particles::Int64,
     ) where {H<:AbstractHook}
@@ -30,6 +32,7 @@ struct EnvHelperSharedProps{H<:AbstractHook}
             agent,
             hook,
             n_steps_before_actions_update,
+            n_actions_updates_per_episode,
             elliptical_a_b_ratio,
             n_particles,
             fill(0, n_particles),

src/RL/Envs/LocalCOMWithAdditionalShapeRewardEnv.jl

@@ -165,7 +165,6 @@ function update_reward!(
     env_helper::LocalCOMWithAdditionalShapeRewardEnvHelper,
     particle::ReCo.Particle,
 )
-    normalization = env_helper.shared.n_particles # TODO: Add factor from steps
     n_neighbours = env_helper.n_neighbours[particle.id]
 
     if n_neighbours == 0
@@ -191,7 +190,7 @@ function update_reward!(
             )
         end
 
-        env.shared.reward = reward / normalization
+        set_normalized_reward!(env, reward, env_helper)
     end
 
     return nothing

src/RL/Envs/OriginEnv.jl

@@ -94,13 +94,10 @@ end
 function update_reward!(
     env::OriginEnv, env_helper::OriginEnvHelper, particle::ReCo.Particle
 )
-    normalization = env_helper.shared.n_particles # TODO: Add factor from steps
-
     reward = minimizing_reward(
         env_helper.distances_to_origin[particle.id], env_helper.max_distance_to_origin
     )
-
-    env.shared.reward = reward / normalization
+    set_normalized_reward!(env, reward, env_helper)
 
     return nothing
 end

src/RL/RL.jl

@@ -70,6 +70,7 @@ function run_rl(;
     packing_ratio::Float64=0.15,
     show_progress::Bool=true,
     reward_discount::Float64=1.0,
+    show_simulation_progress::Bool=true,
 ) where {E<:Env}
     @assert 0.0 <= elliptical_a_b_ratio <= 1.0
     @assert n_episodes > 0
@@ -98,8 +99,15 @@ function run_rl(;
 
     hook = TotalRewardPerEpisode()
 
+    n_actions_updates_per_episode = ceil(Int64, episode_duration / update_actions_at)
     env_helper_shared = EnvHelperSharedProps(
-        env, agent, hook, n_steps_before_actions_update, elliptical_a_b_ratio, n_particles
+        env,
+        agent,
+        hook,
+        n_steps_before_actions_update,
+        n_actions_updates_per_episode,
+        elliptical_a_b_ratio,
+        n_particles,
     )
 
     env_helper_args = (
@@ -127,7 +135,13 @@ function run_rl(;
         agent(PRE_EPISODE_STAGE, env)
 
         # Episode
-        ReCo.run_sim(dir; duration=episode_duration, seed=episode, env_helper=env_helper)
+        ReCo.run_sim(
+            dir;
+            duration=episode_duration,
+            seed=episode,
+            env_helper=env_helper,
+            show_progress=show_simulation_progress,
+        )
 
         env.shared.terminated = true
 

src/RL/Reward.jl

@@ -1,3 +1,16 @@
 function minimizing_reward(value::Float64, max_value::Float64)
     return exp(-0.5 * (value / (max_value / 3))^2)
 end
+
+function reward_normalization(env_helper::EnvHelper)
+    return env_helper.shared.n_particles * env_helper.shared.n_actions_updates_per_episode
+end
+
+function set_normalized_reward!(
+    env::Env, unnormalized_reward::Float64, env_helper::EnvHelper
+)
+    normalization = reward_normalizaion(env_helper)
+    env.shared.reward = unnormalized_reward / normalization
+
+    return nothing
+end