v1.2.0

• Value net trained with EAS bonuses
• Increased value net L1 to 2048
• Quantised value net
• Adjusted EAS params
• Moved movegen library into files to increase performance
• Tuned params of the engine
• Heavily modified contempt code

Estimated Elo: 3150

v1.1.0

This update includes some small improvements for better speed and adjusts parameters to increase overall playing strength. However, these changes might have a negative impact on the EAS score. Because of this, I plan to bring back those parameters in version 2.0.0 to balance things out.

I measured an increase of 120 Elo compared to version 1.0.0, but the improvement in performance when playing against other engines seems to be a bit smaller.

v1.0.0

This engine was developed to harvest some aggressiveness from the MCTS search method. Current estimated Elo is around 3000.

Implemented Features:

• Tree reuse
• LRU
• FPU
• Gini scaling
• Scaling C with search duration
• Variance scaling
• Exploration scaling
• Threats in value net
• Defenses in value net
• Horizontal mirroring in value net
• Softmax policy temperature at root
• Extended time manager
• Multithreading
• Hashtable

Jackal currently supports multiple threads, but due to some unidentified bug it doesn't play better with more than 1 thread. ($100 for fixing PR)

Features to Increase Its EAS:

• Policy bonus to sacrifices
• Filtering positions to only those where sides with lower material won the game
• Asymmetrical contempt in PUCT
• Bonus to positions that have less material than root position

I measured Jackal's EAS to be around 230k, while I also noticed it is very slow to end the games and draws a lot of winning positions. My current guess is that MCTS heavily relies on its neural nets and my current data is just not strong enough to be efficient in end games.

EAS Score	Sacs	Shorts	Bad draws	Avg. win moves	Engine/player
229966	40.00%	26.71%	17.36%	103	Jackal