90fad334b9
More optimizations.
2018-05-15 23:37:35 +02:00
a77c13a0a4
1-ply runs even faster.
2018-05-15 19:29:27 +02:00
260c32d909
oiuhhiu
2018-05-15 18:16:44 +02:00
926a331df0
Some flags from main.py is gone, rolls now allow a face_value of 0 yet
...
again and it is possible to play against the ai. There is no flag
for this yet, so this has to be added.
2018-05-13 23:54:13 +02:00
ba4ef86bb5
Board rep can now be inferred from file after being given once.
...
We can also evaluate multiple times by using the flag "--repeat-eval".
The flag defaults to 1, if not provided.
2018-05-12 12:14:47 +02:00
1aa9cf705f
quack without leaks
2018-05-11 21:24:10 +02:00
93224864a4
More comments, backprop have been somewhat tested in the eager_main.py
...
and normal_main.py.
2018-05-11 13:35:01 +02:00
504308a9af
Yet another input argument, "--ply", 0 for no look-ahead, 1 for a single
...
look-ahead.
2018-05-10 23:22:41 +02:00
3b57c10b5a
Saves calling tf.reduce_mean on all values once.
2018-05-10 22:57:27 +02:00
6131d5b5f4
Added comments for Christoffer!
2018-05-10 19:25:28 +02:00
1aedc23de1
1-ply now works again.
2018-05-10 19:13:18 +02:00
2d84cd5a0b
1-ply now works again.
2018-05-10 19:06:53 +02:00
396d5b036d
All values for boards and all rolls can now be calculated
2018-05-10 18:41:21 +02:00
4efb229d34
Added a lot of comments
2018-05-10 15:28:33 +02:00
f2a67ca92e
All board reps should now work as input.
2018-05-10 10:49:25 +02:00
9cfdd7e2b2
Added a verbosity flag, --verbose, which allows for printing of
...
variables and such.
2018-05-10 10:39:22 +02:00
6429e0732c
We should now be able to both train and eval as per usual.
...
I've added a file "global_step", which works as the new global_step
counter, so we can use it for exp_decay.
2018-05-09 23:15:35 +02:00
cb7e7b519c
Getting closer to functionality. We're capable of evaluating moves
...
and a rework of global_step has begun, such that we now use
episode_count as a way of calculating exp_decay, which have been
implemented as a function.
2018-05-09 22:22:12 +02:00
9a2d87516e
Ongoing rewrite of network to use an eager model. We're now capable of
...
evaluating a list of states with network.py. We can also save and
restore models.
2018-05-09 00:33:05 +02:00
ac6660e05b
Added board-rep as cli argument, to state which input-board-rep to use.
...
Also fixed weird nesting of difference_in_values.
2018-05-06 20:52:35 +02:00
1f8485f54e
No longer use n_ply, shit's too slow man.
...
Added extra logging, now logs the average difference in values
between trainings.
Also fixed bug with the length of quack-norm.
Also added cli argument; use-baseline, if set, the baseline-model
will be used.
2018-05-06 20:41:07 +02:00
1db469709a
make_move now calls n_ply to search deeper and potentially give
...
better moves. It's hella fucking slow.
2018-05-02 01:06:23 +02:00
695a3d43db
Fixed n_ply and actually added a comma in main.py. *clap Christoffer*
2018-05-01 20:39:29 +02:00
c530aa688d
flipidip
2018-05-01 13:48:42 +02:00
3f6849048e
added network_test and some comments
2018-04-29 12:14:14 +02:00
afa6504b05
ply again again
2018-04-26 16:49:49 +02:00
9428a00c11
add "--force-creation" flag to force model creation
2018-04-26 11:43:19 +02:00
48a5f6cbb6
Moved "do_ply" out of "calculate_2_ply", in an effort to be able to
...
eventually do further plies, however some rewriting of the current
"do_ply" will be needed, as described in a comment.
2018-04-26 09:42:03 +02:00
8899c5c2d9
Fixed potential bug in regards to scores in 2-ply calculation.
2018-04-25 00:51:04 +02:00
0509a51fd3
Added baseline model for testing
2018-04-24 22:30:58 +02:00
349ad718f1
Moved gen_21_rolls into the 2-ply method, so it can be correctly used like the good helper method that it is
2018-04-23 00:45:31 +02:00
e5cc54d3e0
Added a normalised version of quack
2018-04-23 00:35:25 +02:00
160f5bd737
added some comments and removed some old code
2018-04-22 19:13:46 +02:00
77d82f6883
Added code for 2-ply look-ahead
2018-04-22 15:07:19 +02:00
1062b72bda
fix typo
2018-04-19 16:04:49 +02:00
66589dfde3
fixed global step, now using exp decay
2018-04-19 16:01:19 +02:00
cba0f67ae2
fixed *the* bug
2018-04-19 15:22:00 +02:00
611f6cdba0
Changed alpha to learning_rate
2018-04-15 23:53:35 +02:00
7d29fc02f2
Added global step + exponential decay
2018-04-14 23:11:20 +02:00
17f5b62e9b
proper Tesauro board representation
2018-03-28 14:36:52 +02:00
fda2c6e08d
parametric board representation in network
2018-03-28 12:00:47 +02:00
abce56dd40
fix typo
2018-03-27 23:13:59 +00:00
95b12a6c35
Added another board_rep
2018-03-28 00:33:39 +02:00
2654006222
fix wrongful mergings
2018-03-27 13:02:36 +02:00
c248ca0452
Merge branch 'fuck_git' into 'rework-1'
...
# Conflicts:
# network.py
2018-03-27 10:15:51 +00:00
f43108c239
Training using slightly revamped version of our own board rep. Not sure if works yet.
2018-03-27 04:06:08 +02:00
006f791727
Functioning network using board representation shamelessly ripped from Tesauro
2018-03-27 02:26:15 +02:00
4c43bf19a3
Add evaluation variance benchmark
...
To do a benchmark for `pubeval`, run `python3 main.py --bench-eval-scores
--eval-methods pubeval`
Logs will be placed in directory `bench`
Use `plot_bench(data_path)` in `plot.py` for plotting
2018-03-26 16:45:26 +02:00
1f1e806306
fix errant whitespace
2018-03-26 15:55:48 +02:00
98c9af72e7
rework network
2018-03-22 15:30:47 +01:00
