feat: add ela_recompute_every parameter for ELA feature recomputation control

Author: Grzmro

das/env/das_env.py

@@ -26,10 +26,6 @@
 from das.env.reward import compute_reward
 from das.optimizers.base import get_checkpoints
 
-# Recompute ELA every ~500 new population samples.  pflacco runs regression,
-# nearest-neighbour search, and IC calculations on every call — running it
-# every step would dominate wall-clock time for long training runs.
-_ELA_RECOMPUTE_THRESHOLD = MAX_HISTORY_SAMPLE // 5
 
 
 class DASEnv(gym.Env):
@@ -68,6 +64,7 @@ def __init__(
         reward_option: int = 1,
         n_individuals: int = 100,
         seed: int | None = None,
+        ela_recompute_every: int = MAX_HISTORY_SAMPLE // 5 # ~500,
     ):
         super().__init__()
         self.problem_ids = problem_ids
@@ -79,6 +76,7 @@ def __init__(
         self.reward_option = reward_option
         self.n_individuals = n_individuals
         self._seed = seed
+        self._ela_recompute_every = max(1, ela_recompute_every)
 
         n_actions = len(optimizers)
         obs_dim = observation_dim(n_actions)
@@ -293,7 +291,10 @@ def _build_observation(self) -> np.ndarray:
         # _ela_cache starts as zeros (correct before 50 samples) and is reset
         # each episode, so stale features from a previous episode never leak in.
         current_len = len(self._x_history) if self._x_history is not None else 0
-        if current_len >= 50 and current_len - self._ela_cache_len >= _ELA_RECOMPUTE_THRESHOLD:
+        if current_len >= 50 and (
+            self._ela_cache_len == 0
+            or current_len - self._ela_cache_len >= self._ela_recompute_every
+        ):
             self._ela_cache = compute_ela_features(self._x_history, self._y_history)
             self._ela_cache_len = current_len
 

das/training/common.py

@@ -31,6 +31,7 @@ def _init():
             reward_option=cfg["reward_option"],
             n_individuals=cfg["n_individuals"],
             seed=cfg.get("seed"),
+            ela_recompute_every=cfg.get("ela_recompute_every", 500),
         )
 
     return _init

das/training/ppo.py

@@ -138,6 +138,7 @@ def run_ppo(args) -> None:
         "reward_option": args.reward_option,
         "n_individuals": args.n_individuals,
         "seed": args.seed,
+        "ela_recompute_every": args.ela_recompute_every,
     }
 
     print(f"Portfolio : {args.portfolio}")

train.py

@@ -75,6 +75,16 @@ def _add_shared_args(p: argparse.ArgumentParser) -> None:
     )
     p.add_argument("--n-individuals", type=int, default=100, help="Population size")
     p.add_argument("--seed", type=int, default=42)
+    p.add_argument(
+        "--ela-recompute-every",
+        type=int,
+        default=500,
+        help=(
+            "Recompute ELA features every N new population samples. "
+            "Set to 1 to recompute on every step (slow but maximally fresh). "
+            "Default: 500."
+        ),
+    )
 
 
 def _parse_args() -> argparse.Namespace: