tweak config

config.lua

@@ -26,17 +26,20 @@ local common_cfg = {
     det_epsilon = false, -- take random actions with probability eps.
     layernorm = false,
 
-    learning_rate = 1.0,
-
     init_zeros = true, -- instead of he_normal noise or whatever.
     graycode = false,
     unperturbed_trial = true, -- do a trial without any noise.
     negate_trials = true, -- try pairs of normal and negated noise directions.
     -- ^ note that this now doubles the effective trials.
     time_inputs = true, -- binary inputs of global frame count
+    normalize_inputs = false,
 
     ars_lips = false,
     adamant = false, -- run steps through AMSgrad.
+    adam_b1 = math.pow(10, -1 / 1), -- fewer trials, more momentum!
+    adam_b2 = math.pow(10, -1 / 50),
+    adam_eps = intmap(-1), -- focus on b1 rather than b2.
+    adam_debias = true,
 
     cap_time = 300,
     timer_loser = 1/2,
@@ -44,21 +47,19 @@ local common_cfg = {
 }
 
 local cfg = {
+    -- best performance so far, albeit a bit slow to train.
+    -- NOTE: uncomment for original values used.
     log_fn = 'log.csv', -- can be nil to disable logging.
     params_fn = nil, -- can be nil to generate based on param count.
-
-    deterministic = true,
-
-    epoch_trials = 20,
-    epoch_top_trials = 10,
-    learning_rate = 1.0,
-
-    ars_lips = true,
-    deviation = 0.1,
-    weight_decay = 0.004,
-
-    cap_time = 300,
     starting_lives = 1,
+    deterministic = true,
+    layernorm = true,
+    epoch_trials = 24, --20,
+    epoch_top_trials = 18, --10,
+    ars_lips = true,
+    deviation = 0.05, --0.1,
+    weight_decay = 0.0002, --0.0004,
+    learning_rate = 2.0, --1.0
 }
 
 -- TODO: so, uhh..