A more classical approach

training/evaluate.py

@@ -94,9 +94,8 @@ def main():
             # Access the underlying HerdingEnv for dispersion calculation
             inner = env.envs[0] if hasattr(env, "envs") else env.venv.envs[0]
             if not inner.penned[:inner.n_sheep].all():
-                ep_dispersion.append(
-                    pairwise_mean(inner.sheep_pos, inner.n_sheep)
-                )
+                _, radius, _ = inner._flock_stats()
+                ep_dispersion.append(radius)
 
             if first_ep and render_mode == "human":
                 pass   # render() is called inside step()
@@ -134,8 +133,8 @@ def main():
           f"  ({sum(successes)}/{args.episodes})")
     print(f"  Time-to-pen     : {mean_ttp:.1f} steps/sheep"
           f"  (successful episodes only)")
-    print(f"  Flock dispersion: {mean_disp:.2f} m"
-          f"  (mean pairwise distance while active)")
+    print(f"  Flock radius    : {mean_disp:.2f} m"
+          f"  (max sheep-to-COM distance while active)")
     print("=" * 50)