feat: bot training configuration file

Cargo.lock

@@ -320,6 +320,7 @@ version = "0.1.0"
 dependencies = [
  "burn",
  "burn-rl",
+ "confy",
  "env_logger 0.10.0",
  "log",
  "pretty_assertions",
@@ -917,7 +918,7 @@ checksum = "fe6d2e5af09e8c8ad56c969f2157a3d4238cebc7c55f0a517728c38f7b200f81"
 dependencies = [
  "serde",
  "termcolor",
- "unicode-width 0.1.14",
+ "unicode-width 0.2.0",
 ]
 
 [[package]]
@@ -964,6 +965,18 @@ dependencies = [
  "crossbeam-utils",
 ]
 
+[[package]]
+name = "confy"
+version = "1.0.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f29222b549d4e3ded127989d523da9e928918d0d0d7f7c1690b439d0d538bae9"
+dependencies = [
+ "directories",
+ "serde",
+ "thiserror 2.0.12",
+ "toml",
+]
+
 [[package]]
 name = "constant_time_eq"
 version = "0.1.5"
@@ -1524,6 +1537,15 @@ dependencies = [
  "subtle",
 ]
 
+[[package]]
+name = "directories"
+version = "6.0.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "16f5094c54661b38d03bd7e50df373292118db60b585c08a411c6d840017fe7d"
+dependencies = [
+ "dirs-sys 0.5.0",
+]
+
 [[package]]
 name = "dirs"
 version = "5.0.1"
@@ -4784,9 +4806,16 @@ dependencies = [
  "serde",
  "serde_spanned",
  "toml_datetime",
+ "toml_write",
  "winnow",
 ]
 
+[[package]]
+name = "toml_write"
+version = "0.1.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "5d99f8c9a7727884afe522e9bd5edbfc91a3312b36a77b5fb8926e4c31a41801"
+
 [[package]]
 name = "torch-sys"
 version = "0.19.0"

bot/Cargo.toml

@@ -23,3 +23,4 @@ env_logger = "0.10"
 burn = { version = "0.17", features = ["ndarray", "autodiff"] }
 burn-rl = { git = "https://github.com/yunjhongwu/burn-rl-examples.git", package = "burn-rl" }
 log = "0.4.20"
+confy = "1.0.0"

bot/src/burnrl/main.rs

@@ -17,38 +17,12 @@ fn main() {
     // let dir_path = &args[2];
 
     let path = format!("bot/models/burnrl_{algo}");
-    let conf = Config {
+    println!(
-        save_path: Some(path.clone()),
+        "info: loading configuration from file {:?}",
-        num_episodes: 30, // 40
+        confy::get_configuration_file_path("trictrac_bot", None).unwrap()
-        max_steps: 1000,  // 1000 max steps by episode
+    );
-        dense_size: 256,  // 128  neural network complexity (default 128)
+    let mut conf: Config = confy::load("trictrac_bot", None).expect("Could not load config");
-
+    conf.save_path = Some(path.clone());
-        gamma: 0.9999, // 0.999 discount factor. Plus élevé = encourage stratégies à long terme
-        tau: 0.0005, // 0.005 soft update rate. Taux de mise à jour du réseau cible. Plus bas = adaptation
-        // plus lente moins sensible aux coups de chance
-        learning_rate: 0.001, // 0.001 taille du pas. Bas : plus lent, haut : risque de ne jamais
-        // converger
-        batch_size: 128, // 32 nombre d'expériences passées sur lesquelles pour calcul de l'erreur moy.
-        clip_grad: 70.0, // 100 limite max de correction à apporter au gradient (default 100)
-
-        // SAC
-        min_probability: 1e-9,
-
-        // DQN
-        eps_start: 0.9, // 0.9  epsilon initial value (0.9 => more exploration)
-        eps_end: 0.05,  // 0.05
-        // eps_decay higher = epsilon decrease slower
-        // used in : epsilon = eps_end + (eps_start - eps_end) * e^(-step / eps_decay);
-        // epsilon is updated at the start of each episode
-        eps_decay: 2000.0, // 1000 ?
-
-        // PPO
-        lambda: 0.95,
-        epsilon_clip: 0.2,
-        critic_weight: 0.5,
-        entropy_weight: 0.01,
-        epochs: 8,
-    };
     println!("{conf}----------");
 
     match algo.as_str() {

bot/src/burnrl/utils.rs

@@ -3,28 +3,39 @@ use burn::nn::Linear;
 use burn::tensor::backend::Backend;
 use burn::tensor::Tensor;
 use burn_rl::base::{Agent, ElemType, Environment};
+use serde::{Deserialize, Serialize};
 
+#[derive(Serialize, Deserialize)]
 pub struct Config {
     pub save_path: Option<String>,
-    pub max_steps: usize,
+    pub max_steps: usize, // max steps by episode
     pub num_episodes: usize,
-    pub dense_size: usize,
+    pub dense_size: usize, // neural network complexity
 
+    // discount factor. Plus élevé = encourage stratégies à long terme
     pub gamma: f32,
+    // soft update rate. Taux de mise à jour du réseau cible. Plus bas = adaptation plus lente moins sensible aux coups de chance
     pub tau: f32,
+    // taille du pas. Bas : plus lent, haut : risque de ne jamais
     pub learning_rate: f32,
+    // nombre d'expériences passées sur lesquelles pour calcul de l'erreur moy.
     pub batch_size: usize,
+    // limite max de correction à apporter au gradient (default 100)
     pub clip_grad: f32,
 
-    // for SAC
+    // ---- for SAC
     pub min_probability: f32,
 
-    // for DQN
+    // ---- for DQN
+    // epsilon initial value (0.9 => more exploration)
     pub eps_start: f64,
     pub eps_end: f64,
+    // eps_decay higher = epsilon decrease slower
+    // used in : epsilon = eps_end + (eps_start - eps_end) * e^(-step / eps_decay);
+    // epsilon is updated at the start of each episode
     pub eps_decay: f64,
 
-    // for PPO
+    // ---- for PPO
     pub lambda: f32,
     pub epsilon_clip: f32,
     pub critic_weight: f32,