jw1912 / akimbo Goto Github PK

Hi Jamie,

I am bit late for reporting ;) but I thought I let you know that I had to compile Akimbo myself
for my old hardware since the 0.4.0 release, while V1 still worked in 0.3.0.

As my current hardware from early 2009 (!) supports just SSE2, SSE3, SSSE3 and SSE4.1 but not popcount, I suspect
that the later releases somehow all are compiled with it. I guess this is not intended for V1 as it defeats
the purpose it. May be you can check again?

Best regards,
Guenther (RWBC XB/UCI chronology)

For better memory locality, handle scores internally by a new score type that contains midgame and endgame scores:

struct Score(i16, i16);

It's sad that it's not possible to compile the newest version on Android and Linux devices.

congrats with the new v0.4.0 ! I managed to compile it (on Linux) and it runs fine.

i'm a bit amazed its evaluation consists only of a set of tapered piece-square tables, tuned using Texel's tuning method .. does this mean Akimbo would play differently when just changing those PSTs ? I guess so .. however, i never did any Texel tuning, i can only imagine how it works, no clue how to generate such alternative values ..

can Texel tuning be compared to training an NN ? it needs many games as input ..

also : could PST values be created which reflect an approach to play a centain opening ?

As a result of originally intending to be a compact engine, code is quite dense and not idiomatic. Aim to refactor to more be more idiomatic as well as integrate the tuner directly into this repository's code.

Large akimbo is unbased, must be made tiny again

Hi there,
i was wondering how i can train a net by myself. I tried to train one and got this done with the bullet trainer. But when i tried to compile the engine with my own net it says the net is too small. It might be helpful if you could show me which parameters you use to train the nets.

Many thanks in advance

And why?

Just to make sure we use the right name.

Gabor Szots
CCRL

I/O needs to have consistent error handling as at the moment its just arbitrary behaviour I decided on.

Example

Sending go depth for depths greater than 128 (which some GUIs do, e.g. go depth 245 in place of go infinite) just returns a depth 1 search.

The evals and hash methods are used only in parse_fen to initialise values, when they could be directly incorporated to save a lot of space (repetition of code between the three methods).

Add chess960 support.