remove some old comments

main.lua

@@ -2,8 +2,6 @@ local globalize = require("strict")
 
 -- configuration.
 
---randomseed(11)
-
 local cfg = require("config")
 local gcfg = require("gameconfig")
 
@@ -93,6 +91,9 @@ local lerp             = util.lerp
 local softchoice       = util.softchoice
 local unperturbed_rank = util.unperturbed_rank
 
+local game = require("smb")
+game.overlay = cfg.enable_overlay
+
 -- utilities.
 
 local log_map = {
@@ -158,12 +159,6 @@ local function make_network(input_size)
     return nn.Model({nn_x, nn_tx}, {nn_z})
 end
 
--- and here we go with the game stuff.
--- which was all refactored out, so this comment looks a little silly now.
-
-local game = require("smb")
-game.overlay = cfg.enable_overlay
-
 -- learning and evaluation.
 
 local function prepare_epoch()
@@ -174,12 +169,6 @@ local function prepare_epoch()
     empty(trial_noise)
     empty(trial_rewards)
 
-    -- TODO: (optionally) save memory.
-    --       generate noise as needed by saving the seed
-    --       (the os.time() as of here) and calling nn.normal() each trial.
-    -- of course this doubles the amount of time we spend generating noise,
-    -- but that's a fair tradeoff for dividing memory used by noise by `epoch_trials`.
-
     local precision = (pow(cfg.deviation, 1/-0.51175585) - 8.68297257) / 1.66484392
     if cfg.graycode then
         print(("chosen precision: %.2f"):format(precision))
@@ -187,8 +176,6 @@ local function prepare_epoch()
 
     for i = 1, cfg.epoch_trials do
         local noise = nn.zeros(#base_params)
-        -- NOTE: change in implementation: deviation is multiplied here
-        --       and ONLY here now.
         if cfg.graycode then
             --local precision = 1 / cfg.deviation
             --print(cfg.deviation, precision)
@@ -198,8 +185,6 @@ local function prepare_epoch()
             for j = 1, #base_params do
                 noise[j] = nn.uniform() < 0.5 and noise[j] or -noise[j]
             end
-            -- TODO? wrap/bound domain to [-1,1].
-            --       dunno if this will work with the learning rate stuff.
         else
             for j = 1, #base_params do
                 noise[j] = cfg.deviation * nn.normal()
@@ -334,7 +319,6 @@ local function learn_from_epoch()
         for i, v in ipairs(top_rewards) do top_rewards[i] = v / reward_dev end
     end
 
-    -- NOTE: step no longer directly incorporates learning_rate.
     for i = 1, cfg.epoch_trials do
         local ind = (i - 1) * 2 + 1
         local pos = top_rewards[ind + 0]