Upload folder using huggingface_hub

code/train_value_iteration.py

@@ -186,12 +186,20 @@ def main():
                    help="States are sampled from walks of 1..this many moves. Past "
                         "~20 every state is fully mixed (God's Number), so a larger "
                         "value buys nothing.")
-    p.add_argument("--update-threshold", type=float, default=0.05,
+    p.add_argument("--update-threshold", type=float, default=0.01,
                    help="Refresh the target network once training loss falls below "
-                        "this. Lower is more stable and slower.")
-    p.add_argument("--min-update-steps", type=int, default=500,
-                   help="Never refresh the target network more often than this, so "
-                        "an early lucky batch cannot ratchet it forward.")
+                        "this. This is meant to be the binding constraint: it lets "
+                        "an easy early target move on fast and gives a harder later "
+                        "one as long as it needs. Set it above the loss the net "
+                        "actually reaches and it never binds, leaving --min-update-"
+                        "steps to pin the cadence -- measured at 0.05, the net sat "
+                        "at 0.0004 and idled ~800 of every 1000 steps.")
+    p.add_argument("--min-update-steps", type=int, default=200,
+                   help="Floor on the refresh interval, so one lucky batch cannot "
+                        "ratchet the target forward. Each refresh propagates "
+                        "accuracy roughly one move further from solved, so this "
+                        "floor times ~30 is the run length; measured re-convergence "
+                        "after a refresh is ~200 steps.")
     p.add_argument("--eval-every", type=int, default=2000)
     p.add_argument("--eval-n", type=int, default=200)
     p.add_argument("--eval-depths", default="3,6,10,15,20")