-
Notifications
You must be signed in to change notification settings - Fork 2
/
TODO
74 lines (70 loc) · 3 KB
/
TODO
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
TODO
------------------
Now: (current focus)
add debug levels
change uct_* params to tree_*
option to output games to sgf with explanations (see gomill-savesgf)
add random chance of not using a playout heuristic (per heuristic)
weighted "random" ordering of playout heuristics
add an auto-tuning version, like Mogo now has
solve problem of LGRF hindering ladder-handling
reduce Engine.h includes, to optimize build process
move Parameter defaults and inits out of Engine
make time_k boardsize independant
add kgs chat command with winrate
don't check for opening book moves once we are out of the book
Soon: (before next major release)
larger patterns
add results from first playout moves to tree
try UCB1-TUNED formula (p.216(169) in SylvainGellyThesis)
prefer games that end sooner? (don't drag on at end)
reduce thinking time if surewin?
think for longer if needed
keep track of board hashes incrementally
speedup features in playouts
setting to use a total number of playouts per game for time testing
patterns only near last move?
fillboard not on 1st/2nd line?
make sure scoring at end of kgs ranked game works (opponent can't "cheat")
tune fillboard's N
look into transposition table-like structure
look at Feasibility for 19x19 (Baier Hendrik), possibly as a feature
look at Dynamic Komi for 19x19 (Baier Hendrik)
try LGRF-1/2 with a variable queue (see 6.3.5 Multiple Reply LGR in Baier Hendrik)
try make progressive bias decay with parent playouts
look why bs=0 even when r2 is high (possibly fixed since stop_early)
look at altering playout_random_chance with extreme winrate
look at a reverse LGR-kinda thing in the tree and playouts
convert feature patterns to ASCII pics to aid visualisation
compare profiling data to that of other programs such as libEGO
try a "step-wise" RAVE (decay value at each capture)
add parameters to web interface
improve the overall web interface look and reduce flickering
option to output SGFs for playouts
Later: (at some point)
tune parameters
add handicap settings
add timestamp to log?
replace FunctionList and ConstantList with std::list<>
replace the Gtp instance variable using boost::bind()
rename/refactor Engine to better name?
move Go:: objects into own files?
move main() into Oakfoam.cc
add prefix to custom commands?
do sanity checks on parameters
use rotational symmetry?
copy all board data with a copy() (groups and liberties)
extensive comparison of playouts with strength
more regression testing
add loadsgf command
setting to filter boardsize, time_settings, etc. for kgs
try a "global RAVE"
try use parent info in children (or grand parent)
make atari-aware playouts dependant on size of group?
make self-atari feature dependant on group size
add hg tracked cgos,kgs?,gomill? scripts
make SL page
support terminal jigo in tree?
add seg fault and exception handling
try use Benson algo in tree and playouts?
progressively add more reward for points diff if winrate very good?