UCI chess engine: Jackal
Author - Tomasz Jaworski
Jackal 2.0.0 what's new?
Overview
This version is a full rewrite from scratch. Pretty much nothing from 1.2.0 survived. Old repository is archived, 2.0.0 lives in a new one.
Estimated elo: 3485 (around 380 elo more than last version)
EAS: 403.35k
Networks
- Networks got a lot bigger — both value and policy
- Value network now uses threat inputs
- Architecture: 82160 (Threat Inputs) → 4096 → 16 → 128 → 3
- Policy architecture was completely changed
- Architecture: 3072 → 8192 → 7840 (Move Index)
- Output buckets based on SEE result
- 7 networks total (3 value + 4 policy) — base + finetuned versions
- Finetuned on filtered aggressive/sacrificial positions, used based on position score
- Engine parameters tuned with SPSA
Search
A lot of things changed here. Some highlights:
- Sac selection bonus - sac moves get a bonus when picking the reported move, scales with material and winning probability
- Sac exploration bonus - sac moves get extra exploration weight during MCTS selection, scales with win probability
- Draw pessimism - draws are penalized during simulation, mostly in opening and middlegame
- Proof number search - one-sided PNS, biases search toward lines closest to being proven when winning
- Butterfly history - added as bonus during search
- Progressive widening - focus search on promising moves early
- Multithreading fix - there was a bug, now it's fixed
Time Manager
Completely new and custom time manager built from scratch. Some of what it includes: soft and hard time limits, phase-based scaling, instability extensions, falling eval extension, visit distribution scaling, sacrifice extension, and more.

Comments
Post a Comment