Skip to content
Xiangqi AI Knowledge
Xiangqi AI basics: engines, interfaces and draw rates.
Basic, nontechnical reading. For technical topics, see “Technical reference”.
39 questions
帥Basics
- What is an "engine"?An engine is a program that calculates and analyzes positions. Pikafish is an open-source, free Xiangqi engine.→
- What is a "GUI"?A GUI is software that lets you talk to an engine conveniently. The engine is the motor and the GUI is the car.→
- What is an "opening book"?An opening book is like a human "game record" that software can follow to save time and lose less.→
- What is the "cloud database"?The cloud database is a free online Xiangqi database with a huge number of positions scored by servers.→
- What is an "endgame tablebase"?An endgame tablebase exhausts every position of certain material sets and records win/draw/loss and winning moves. Barring bugs, it is absolutely accurate.→
- What is an "evaluation library" (审局库)?An evaluation library is an endgame tablebase that records only win/loss/draw results, with no moves, so it is much smaller.→
- What is "NNUE"?NNUE is a neural network architecture suited to CPU inference. It only evaluates positions; the search is still traditional ab pruning.→
- What is a "traditional engine"?A traditional engine uses hand-written evaluation instead of a neural network. It searches faster but has a much lower strength ceiling than NNUE engines.→
- What is "evaluation"?Evaluation scores a position from parameters alone, without calculation. Before NNUE it was hand-written; now a neural network does it.→
- What is "AI" (artificial intelligence)?It depends on the definition. By most people's understanding, any top engine with a neural network counts as AI, Pikafish included.→
- What is FEN?FEN records a Xiangqi position and the side to move as a string of characters, and engines receive positions through it.→
車Hardware & speed
- What are cores (e.g. 128 cores, 256 cores)?A core is a physical CPU core, a hardware spec; more and faster cores mean faster engine calculation.→
- What are instruction sets (such as bmi2 and avx2)?Instruction sets set how efficiently the engine talks to the CPU and affect only speed; the universal build now auto-selects.→
- What is NPS (the "K value")? Does it represent playing strength?NPS is nodes searched per second; it compares machines only and does not show engine strength.→
- What is depth (plies)? Does a fast-climbing depth mean strength?Depth is just the number of search iterations, not moves calculated; how fast it climbs does not show strength.→
- What is the transposition table (hash), and how much should I set?It stores and reuses results of searched positions; a few GB is recommended, and more barely matters.→
- Is a machine with more cores or threads always better?Usually yes, but architecture, frequency and instruction set matter too; judge hardware by nps.→
- How much does computing power matter for a Xiangqi engine?Hard to see in real games, where draws dominate; very clear in unbalanced paired-game tests.→
- What is the difference between multi-core multithreading and a single core?With equal total compute, multithreading is weaker than one thread with more time in most positions, and it is random.→
- Is a GPU useful for Xiangqi engines?Depends on the architecture; Pikafish uses only the CPU, so a graphics card will not make it stronger.→
- What is bench, and what do its results tell you?bench is Pikafish's speed test; it reports total time, nodes and nps so you can compare machines.→
炮Strength & testing
- Which is stronger, Pikafish or commercial paid engines?Pikafish is the strongest Xiangqi engine; "paid is always stronger than free" is false.→
- What is the draw rate of top Xiangqi engines today?Strong engines draw extremely often from balanced openings; every test below is above 94%.→
- EloElo reflects true level through win rates; a rating gap maps to an expected score, and different systems are not comparable.→
- How do you test engines scientifically?Same conditions, varied openings in paired games, enough games with statistical tools; use unbalanced openings when strengths are close.→
- Is it reliable to judge which engine is stronger by "feel" without testing?No. Judging by a few positions or score levels is one-sided; large-sample testing is needed.→
- PtnmlPtnml merges the two paired games of one opening into five outcomes, and is very useful for unbalanced-opening tests.→
- Why do engines give different scores, even different versions of the same engine?Engine scores have no common standard, so scores of different engines or versions cannot be compared directly.→
- Does a stronger engine always play better than a weaker one in every position?No. Stronger means better in more positions, not in every position.→
- Can engine-versus-engine games only end in draws?No. Strong engines draw extremely often, but there is always some chance of a decisive game, even under identical conditions.→
- Why can't engines solve certain positions?Engines are not all-powerful; they may miss the solution in puzzles and advantageous openings, and misjudgments are not rare.→
- Pikafish test dataPikafish 20240314 self-play - draw rate 98.2% from balanced openings, 58.54% from Red-advantage openings.→
馬Xiangqi & AI
- Has Xiangqi been solved?No, not even close. The state count is nearly 40 digits, far beyond any storage or computing power.→
- Do engines calculate using "brute force"?No. Engines prune away most moves with alpha-beta search; being fast does not mean brute force.→
- Is there an "AlphaGo" for Xiangqi?No, but px0 uses the same algorithm. For now, CPU engines are still stronger than GPU engines in Xiangqi.→
- Can large language models (like ChatGPT and DeepSeek) play Xiangqi?No. LLMs don't know the rules and often play illegal moves. Use a dedicated Xiangqi engine instead.→
- What is the "self-learning" of a neural network?A mainstream engine's neural network is trained offline by its author and then released. The engine does not learn while you use it.→
相Rules
- Why are Xiangqi rules controversial?Xiangqi's repetition rules have never been unified, so engines and platforms often disagree. Pikafish defined its own program rules based on the World Xiangqi Rules.→
- What is adjudicated auto-draw?The Pikafish testing platform only auto-draws when the material makes a decisive result impossible, keeping results rigorous.→