top

IN REPO

'lozza.js'

'datagen.js'
'filter.js'
'trainer.js'

NOT IN REPO

'data/weights_sqrrelu_128_5_180_v44_5.js'

'lozzacand.js'
'lozzarepo.js'

'lozza4.js'
'lozza3.js'

--{{{  stuff to try

don't copy acc. move update vars to nodes. then do -update to get back to orig acc.
soft/hard time bounds
try iir if no hash move after iid?
try no iid and no iir
perspective! engine easy - can't figure out trainer
take pv out of tt
train 2 nets with queens and without
root sort
rank killers before captures
try killers in qs (bad takes)
add centre to noisy ranking to stop lots of same scores e.g. NxP
somehow merge pst eval with nnue eval. i.e. pst eval handles low hanging fruit - how?
don't use rank vec, use piece itself
try ignore beta when beta pruning
ditto alpha with alpha pruning and futility
quicksee in slide ranking
output buckets based on game phase or numpieces (available) or mid/end game etc
corrhist
conthist
improving - done but can't get it to pass
lmp - tell move iterator - quicker
test removing or simplifying lonepawns (do in position)
remove king from isattached etc - dont capture king - castle conditionals a la cwtch - tried but not working
swapping move list psts for move ranking based on queens on board or not - poss in position
try and get rid of MATERIAL - only used in QS prune now
move drawn stuff from eval into isdraw
full width window searches based on movetime not constant - e.g. log?
push harry

--}}}