add presets to config, update config in general

config.lua

@@ -1,3 +1,5 @@
+local preset = 'snes'
+
 local function approx_cossim(dim)
     return math.pow(1.521 * dim - 0.521, -0.5026)
 end
@@ -13,75 +15,107 @@ local function intmap(x)
 end
 
 local common_cfg = {
+    -- read-only modes:
     playable_mode = false,
     playback_mode = false,
+
+    -- controlling in-game mechanics:
+    starting_world = 1, -- set to 0 to randomize the world every epoch.
+    starting_level = 1, -- set to 0 to randomize the level every epoch.
     start_big = false,
     starting_lives = 0,
+    min_time = 100,
+    max_time = 300,
+    timer_loser = 1/2, -- decrement the timer when mario is slacking off.
 
+    -- rewards:
+    decrement_reward = false, -- bad idea, encourages mario to run into goombas.
+    score_multiplier = 1, -- how much the ingame score influences our rewards.
+
+    -- network initialization:
+    init_zeros = false, -- instead of random normal noise.
+
+    -- network inputs (connections):
+    time_inputs = true, -- binary inputs of global frame count
+
+    -- network layers:
+    layernorm = false,
+
+    -- network evaluation (sampling joypad):
     frameskip = 4,
     -- true greedy epsilon has both deterministic and det_epsilon set.
     deterministic = false, -- use argmax on outputs instead of random sampling.
     det_epsilon = false, -- take random actions with probability eps.
-    layernorm = false,
 
-    init_zeros = true, -- instead of he_normal noise or whatever.
-    graycode = false,
-    unperturbed_trial = true, -- do a trial without any noise.
+    -- evolution strategy and non-rate hyperparemeters:
+    es = 'ars',
+    ars_lips = false, -- for ARS.
+    epoch_top_trials = 9999, -- for ARS.
+
+    -- sampling:
+    unperturbed_trial = true, -- perform an extra trial without any noise.
+    graycode = false, -- for ARS.
     negate_trials = true, -- try pairs of normal and negated noise directions.
     -- AKA antithetic sampling. note that this doubles the number of trials.
-    time_inputs = true, -- binary inputs of global frame count
-    normalize_inputs = false,
+    min_refresh = 0.1, -- for SNES.
 
+    -- epoch-related rates:
     learning_rate = 1.0,
-    mean_adapt = 1.0, -- for xNES
+    mean_adapt = 1.0, -- for xNES.
     weight_decay = 0.0,
-    sigma_decay = 0.0,
-    min_refresh = 0.2,
-
-    es = 'ars',
-    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,
-
-    min_time = 100,
-    max_time = 300,
-    timer_loser = 1/2,
-    decrement_reward = false, -- bad idea, encourages mario to run into goombas.
-    score_multiplier = 1, -- how much the ingame score influences our rewards.
-
-    starting_world = 1, -- set to 0 for random!
-    starting_level = 1, -- set to 0 for random!
+    sigma_decay = 0.0, -- for SNES.
 }
 
-local cfg = {
-    log_fn = 'logs-snes.csv',
-    params_fn = 'params-snes.txt',
+local cfg
+if preset == 'snes' then
 
-    decrement_reward = true,
-    score_multiplier = 5,
+    cfg = {
+        es = 'snes',
 
-    starting_world = 0,
-    starting_level = 1,
-    --starting_lives = 1,
-    min_time = 300,
-    max_time = 300,
+        log_fn = 'logs-snes.csv',
+        params_fn = 'params-snes.txt',
 
-    deterministic = false, --true,
+        starting_lives = 1,
+        min_time = 300,
+        timer_loser = 1.0,
 
-    epoch_trials = 32,
-    epoch_top_trials = 9999,
-    negate_trials = true,
+        epoch_trials = 100,
+        negate_trials = false,
 
-    es = 'snes',
-    learning_rate = 0.5,
-    mean_adapt = 0.5,
-    deviation = 0.5,
-    weight_decay = 0.025,
-    sigma_decay = 0.001,
-}
+        min_refresh = 0.2,
+
+        learning_rate = 0.1, -- TODO: rename to learn_primary or something.
+        mean_adapt = 0.5, -- TODO: rename to learn_secondary or something.
+        deviation = 1.0,
+        weight_decay = 0.01, -- note: multiplied by its std, and mean_adapt.
+        sigma_decay = 0.01, -- note: multiplied by learning_rate.
+    }
+
+elseif preset == 'ars' then
+
+    cfg = {
+        es = 'ars',
+        epoch_top_trials = 20,
+        ars_lips = false,
+
+        log_fn = 'logs-ars.csv',
+        params_fn = 'params-ars.txt',
+
+        min_time = 300,
+        timer_loser = 1.0,
+
+        epoch_trials = 25,
+
+        learning_rate = 1.0,
+        deviation = 0.1,
+        weight_decay = 0.0025,
+    }
+
+else
+
+    error("invalid preset: "..tostring(preset))
+
+end
 
 -- TODO: so, uhh..
 -- what happens when playback_mode is true but unperturbed_trial is false?
@@ -108,8 +142,7 @@ assert(not cfg.ars_lips or cfg.unperturbed_trial,
        "cfg.unperturbed_trial must be true to use cfg.ars_lips")
 assert(not cfg.ars_lips or cfg.negate_trials,
        "cfg.negate_trials must be true to use cfg.ars_lips")
-
-assert(not cfg.adamant,
-       "cfg.adamant not yet re-implemented")
+assert(not (cfg.es == 'snes' and cfg.negate_trials),
+       "cfg.negate_trials is not yet compatible with SNES")
 
 return cfg