Duffish is a highly aggressive chess variant engine derived from Fairy‑Stockfish. It inherits full support for Xiangqi while introducing dedicated evaluation features that promote sharp, sacrificial, and uncompromising play.
Aggressiveness in Duffish is quantified using the Xiangqi Engine Aggressiveness Score (XEAS), a metric developed for Xiangqi engines.
Typical engines score between 30,000 and 80,000 XEAS. Duffish 3 targets a comparable level of aggression -around 134,597 -in the Xiangqi domain while remaining highly competitive in playing strength.
On the Kaka's rating list, Duffish 1 reaches approximately 3060 Elo, well above human grandmaster level.
| Engine | Games | XEAS | Elo |
|---|---|---|---|
| Duffish 3 | 500 | 134,597 | 3060.3 |
| Pikafish HCE 20240522 | 500 | 72,639 | 3107.8 |
Duffish supports all standard UCI options from Fairy‑Stockfish, plus the following custom parameters that fine‑tune its aggressive behavior.
| Option | Range | Default | Description |
|---|---|---|---|
| Aggressiveness | 0-300 | 165 | Controls the engine's overall attacking style. >100 = more aggressive: tighter pruning margins, stronger attacking move preference, easier attack/sacrifice continuation handling, higher king-attack pressure weight, and amplified evaluation. <100 = more conservative. 100 = neutral. |
| DrawValue | -100-100 | -39 | Sets the draw score for the side to move at the root only; the opponent's draw score is unchanged. Under the AXF chasing rule in Xiangqi, this effect may be partially suppressed, but it remains active in other supported variants. |
| SacBonus | 0-200 | 45 | Adds an evaluation bonus when the root side has sacrificed material and the position shows enough attacking compensation. Higher values encourage sacrificial play. |
| DynamicComp | 0-150 | 63 | Adds dynamic compensation based on attack quality and evaluation trends, helping the engine avoid overly passive play in promising attacking positions. |
| AdvisorBreakBonus | 0-150 | 35 | Adds a bonus when the opponent has lost advisors, encouraging attacks that damage the king's defensive structure. |
| BishopBreakBonus | 0-150 | 16 | Adds a bonus when the opponent has lost bishops/elephants, encouraging attacks that weaken defensive coverage and open the board. |
| MatScale | 0-150 | 10 | Scales positive evaluations upward based on remaining material and sacrifice size, making favorable attacking/material-compensation positions score more clearly. |
| SacDetect | 0-150 | 12 | Enables and weights Patricia-style sacrifice detection from the root side's perspective. Higher values increase bonuses for sacrifice sequences with sufficient attack pressure, trace, and compensation quality. |
| DrawMatBias | -100-100 | 25 | Biases draw evaluations according to root-side material imbalance, making drawn lines slightly more or less attractive depending on the material situation. |
| EvalDecay | 0-100 | 35 | Gradually reduces non-zero evaluations as the move-rule counter increases, dampening optimistic scores in positions approaching draw-rule territory. |
Duffish builds exactly like Fairy‑Stockfish. Please refer to the Fairy‑Stockfish build instructions for your platform.
A quick start:
git clone https://github.com/lxsgx23/Duffish.git
cd Duffish/src
make -j profile-buildReplace profile-build with the appropriate target if necessary (e.g., build for a debug version).
Duffish is a standard UCI engine and works with any UCI‑compatible GUI (e.g. Sharkchess).
Duffish is built on Fairy‑Stockfish by Fabian Fichter, which in turn derives from Stockfish.
The aggressiveness scoring methodology is inspired by the Patricia and the Engine Aggressiveness Score (EAS) concept.
Thanks x , rui for testing my engine!
Duffish is distributed under the GNU General Public License v3.0.