c3929d8aa1
remove some ancient useless code
2018-06-16 00:55:41 +02:00
2fe009b5fe
don't base max_time on number of trials (fixes playback mode)
2018-06-16 00:55:41 +02:00
b0058db80e
TODOs and delete an unused variable
2018-06-16 00:38:09 +02:00
155f868f56
another attempt at fixing preset argument
2018-06-16 00:37:19 +02:00
e3695bfb84
rename weight* to param* outside of nn.lua
2018-06-16 00:33:47 +02:00
f3fc95404c
overhaul learning rates:
...
- rename mean_adapt to weight_rate
- sigma and covar update rates can be specified separately
(sigma_rate, covar_rate)
- base decays on current rates instead of initially configured rates
(this might break stuff)
- base_rate takes the place of learning_rate
2018-06-16 00:29:15 +02:00
f512f8ac3a
add sigma decay to xNES
2018-06-14 22:40:39 +02:00
63583789c3
use locals; fix fitness_shaping and graycode
2018-06-13 22:52:37 +02:00
1fba61e1b9
one more attempt at fixing the preset argument
2018-06-13 22:46:31 +02:00
6498b4143f
tweak inputs: add power-up status, remove top/bottom tile rows
2018-06-13 20:18:10 +02:00
6fa042eda5
fix preset failing to default
2018-06-13 06:08:32 +02:00
7800510d1f
add xNES preset, add options, allow preset specified by argument
2018-06-13 03:01:54 +02:00
403127bd66
log decisions counter
2018-06-13 03:00:05 +02:00
b4e49d08b9
restore step logging, remove adamant (for now)
2018-06-13 01:42:36 +02:00
5c64fcf395
overhaul SNES (importance sampling, adaptation sampling, etc)
2018-06-13 01:19:32 +02:00
fca4779e56
fix training without an unperturbed trial
2018-06-13 01:00:15 +02:00
b7a9360d6d
add min_time setting (cap_time -> max_time)
2018-06-13 00:59:36 +02:00
ccce6a2d55
sigma tweaks
2018-06-12 05:39:22 +02:00
0d28db0fc4
allow division of input size in Dense layers
2018-06-12 05:37:35 +02:00
50a7ba78f9
make filenames local to main
2018-06-12 05:36:24 +02:00
fa0287d966
add sigma decay; move printing to start of epoch
2018-06-10 19:34:17 +02:00
56f7c01256
fix network loading
2018-06-10 19:34:06 +02:00
bc655979af
display decisions made instead of frame count
2018-06-10 16:48:02 +02:00
19cd10382f
use experimental config/network
2018-06-10 16:41:45 +02:00
401effbc23
insignificant tweaks
2018-06-10 16:41:32 +02:00
3eebbc534a
add SNES optimizer
2018-06-10 16:40:20 +02:00
d87b8e7118
add mean adaptation hyperparameter
2018-06-10 16:38:25 +02:00
47eb173dac
add exists utility function
2018-06-10 16:36:15 +02:00
0100934ac4
add antithetic sampling for xNES
2018-06-10 16:33:38 +02:00
d6cc49cde1
fix learning without negate_trials
2018-06-09 18:56:10 +02:00
bcb6cb9da1
add xNES optimizer
2018-06-09 18:56:10 +02:00
fe9494b0d5
refactor ARS out of main (breaks a bunch of stuff)
2018-06-09 18:56:10 +02:00
d3e6441c40
reduce tile input to 5 per row using new layers
2018-06-09 16:20:20 +02:00
dd5ec3dbde
make network linear
2018-06-09 16:20:07 +02:00
cbb094adc9
restore flagpole bonus, add missing overlay check
2018-06-09 04:35:09 +02:00
9fb98d3fe0
allow setting of world-level, plus random option
2018-06-09 04:34:21 +02:00
81d6b509d0
detect when mario is controllable
2018-06-09 01:43:22 +02:00
9b23327df4
add score multiplier
2018-06-08 23:59:43 +02:00
fec148fb79
don't turbo in playable mode, note overlay bug
2018-06-08 14:51:17 +02:00
c30f07f407
prevent reward gained from fireworks
2018-06-08 14:12:21 +02:00
c40e1f929d
fix skipped inputs on lag frames
2018-06-08 13:48:59 +02:00
e24c3d31a4
use argsort
2018-06-08 13:46:38 +02:00
374fa4d876
cleanup
2018-06-08 02:45:07 +02:00
37d404e77d
reduce embed layer to values actually used ingame
2018-06-07 22:40:31 +02:00
3030e83d00
refactor learn_from_epoch
2018-05-14 01:34:08 +02:00
15f0292485
remove defer_prints option (now always true)
2018-05-12 22:56:04 +02:00
0fb3b1780f
remove some old comments
2018-05-12 22:55:04 +02:00
a836314b8b
refactor game and utility functions
2018-05-12 22:44:53 +02:00
7db43038ac
adjust range of timed inputs to stdev of roughly 1
2018-05-07 16:27:51 +02:00
7357c8ed62
move LayerNorm after Relu
2018-05-07 16:22:48 +02:00
