Dear lichess.org community,
we are excited to announce that release 0.8.0 of the neural network engine CrazyAra 0.8.0 is now published.
Release 0.8.0 comes with a lot of new features and is expected to be the last release which only supports crazyhouse.
The features include:
* Native TensorRT backend
* Multi-GPU support
* Low precision inference (float16, int8)
* MCTS solver (detects forced mating lines)
* Improved UCI interface
* Random root exploration
* Tablebase support (not used for crazyhouse)
* Full OS support (binary for Linux, Windows and Mac)
All details can be found in the release note:
github.com/QueensGambit/CrazyAra/releases/tag/0.8.0
Our previous paper is also now published on frontiers with open access:
www.frontiersin.org/articles/10.3389/frai.2020.00024/full
@CrazyAra will be hosted for 1.5 days starting from Sunday, 17 May 2020 10:00 (UTC):
www.timeanddate.com/countdown/generic?iso=20200517T12&p0=992&msg=Countdown%3A+CrazyAra+on+lichess.org+-+Release+0.8.0+&font=cursive
For the first 8 hours, CrazyAra will be running on a dual GPU setup (GTX1080ti + RTX2070 OC, float16).
A single RTX2070 OC GPU will be used for the remaining time.
You can challenge CrazyAra to a crazyhouse match in bullet, blitz or classical time control format.
However, you need to set at least one second increment due to additional time delays for lichess-bots.
CrazyAra will be using Model-OS-96, the final model after ~ 2.37 million self-play games.
Detailed information about how this model was trained is available in the master thesis:
Deep Reinforcement Learning for Crazyhouse
ml-research.github.io/papers/czech2019deep.pdf
CrazyAra was initially started as a semester project and exclusively learned to play crazyhouse based on lichess games.
It was first launched on lichess.org on September 7, 2018:
lichess.org/forum/general-chess-discussion/crazyhouse-ai-crazyara
using a full python backend.
Good luck at your games and we hope you enjoy the new release,
~IQ_QI (Johannes Czech)
we are excited to announce that release 0.8.0 of the neural network engine CrazyAra 0.8.0 is now published.
Release 0.8.0 comes with a lot of new features and is expected to be the last release which only supports crazyhouse.
The features include:
* Native TensorRT backend
* Multi-GPU support
* Low precision inference (float16, int8)
* MCTS solver (detects forced mating lines)
* Improved UCI interface
* Random root exploration
* Tablebase support (not used for crazyhouse)
* Full OS support (binary for Linux, Windows and Mac)
All details can be found in the release note:
github.com/QueensGambit/CrazyAra/releases/tag/0.8.0
Our previous paper is also now published on frontiers with open access:
www.frontiersin.org/articles/10.3389/frai.2020.00024/full
@CrazyAra will be hosted for 1.5 days starting from Sunday, 17 May 2020 10:00 (UTC):
www.timeanddate.com/countdown/generic?iso=20200517T12&p0=992&msg=Countdown%3A+CrazyAra+on+lichess.org+-+Release+0.8.0+&font=cursive
For the first 8 hours, CrazyAra will be running on a dual GPU setup (GTX1080ti + RTX2070 OC, float16).
A single RTX2070 OC GPU will be used for the remaining time.
You can challenge CrazyAra to a crazyhouse match in bullet, blitz or classical time control format.
However, you need to set at least one second increment due to additional time delays for lichess-bots.
CrazyAra will be using Model-OS-96, the final model after ~ 2.37 million self-play games.
Detailed information about how this model was trained is available in the master thesis:
Deep Reinforcement Learning for Crazyhouse
ml-research.github.io/papers/czech2019deep.pdf
CrazyAra was initially started as a semester project and exclusively learned to play crazyhouse based on lichess games.
It was first launched on lichess.org on September 7, 2018:
lichess.org/forum/general-chess-discussion/crazyhouse-ai-crazyara
using a full python backend.
Good luck at your games and we hope you enjoy the new release,
~IQ_QI (Johannes Czech)