notwa/smbot
1
0
Fork 0
Code Issues Pull requests Releases Wiki Activity

remove some old comments

Browse source
This commit is contained in:
Connor Olding 2018-05-12 22:55:04 +02:00
parent a836314b8b
commit 0fb3b1780f
1 changed files with 3 additions and 19 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

22
main.lua
View file

@ -2,8 +2,6 @@ local globalize = require("strict")
-- configuration. -- configuration.
--randomseed(11)
local cfg = require("config") local cfg = require("config")
local gcfg = require("gameconfig") local gcfg = require("gameconfig")
@ -93,6 +91,9 @@ local lerp = util.lerp
local softchoice = util.softchoice local softchoice = util.softchoice
local unperturbed_rank = util.unperturbed_rank local unperturbed_rank = util.unperturbed_rank
local game = require("smb")
game.overlay = cfg.enable_overlay
-- utilities. -- utilities.
local log_map = { local log_map = {
@ -158,12 +159,6 @@ local function make_network(input_size)
return nn.Model({nn_x, nn_tx}, {nn_z}) return nn.Model({nn_x, nn_tx}, {nn_z})
end 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. -- learning and evaluation.
local function prepare_epoch() local function prepare_epoch()
@ -174,12 +169,6 @@ local function prepare_epoch()
empty(trial_noise) empty(trial_noise)
empty(trial_rewards) 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 local precision = (pow(cfg.deviation, 1/-0.51175585) - 8.68297257) / 1.66484392
if cfg.graycode then if cfg.graycode then
print(("chosen precision: %.2f"):format(precision)) print(("chosen precision: %.2f"):format(precision))
@ -187,8 +176,6 @@ local function prepare_epoch()
for i = 1, cfg.epoch_trials do for i = 1, cfg.epoch_trials do
local noise = nn.zeros(#base_params) local noise = nn.zeros(#base_params)
-- NOTE: change in implementation: deviation is multiplied here
-- and ONLY here now.
if cfg.graycode then if cfg.graycode then
--local precision = 1 / cfg.deviation --local precision = 1 / cfg.deviation
--print(cfg.deviation, precision) --print(cfg.deviation, precision)
@ -198,8 +185,6 @@ local function prepare_epoch()
for j = 1, #base_params do for j = 1, #base_params do
noise[j] = nn.uniform() < 0.5 and noise[j] or -noise[j] noise[j] = nn.uniform() < 0.5 and noise[j] or -noise[j]
end end
-- TODO? wrap/bound domain to [-1,1].
-- dunno if this will work with the learning rate stuff.
else else
for j = 1, #base_params do for j = 1, #base_params do
noise[j] = cfg.deviation * nn.normal() 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 for i, v in ipairs(top_rewards) do top_rewards[i] = v / reward_dev end
end end
-- NOTE: step no longer directly incorporates learning_rate.
for i = 1, cfg.epoch_trials do for i = 1, cfg.epoch_trials do
local ind = (i - 1) * 2 + 1 local ind = (i - 1) * 2 + 1
local pos = top_rewards[ind + 0] local pos = top_rewards[ind + 0]