Hong Kong Mahjong AI Opponents

How the Hong Kong Mahjong AI Learns

Each AI difficulty level was trained on real game replays from its corresponding mode. The Beginner AI watched beginner games; the Advanced AI learned from advanced games with a 3+ faan minimum. No gameplay data was mixed between modes, so each model genuinely understands the strategy appropriate to its level — not a watered-down version of something harder.

There are no hard-coded rules telling the AI which tile to discard or whether to claim. It learned entirely by watching real games, which means it plays mahjong the way a human who has seen thousands of games would play it.

The models are continuously retrained as more gameplay data accumulates, so the AI keeps improving over time. The opponent you face today is sharper than the one players faced at launch — and it will keep getting better.

What the Mahjong AI Sees at the Table

Before making any decision, the AI takes a snapshot of the entire game state. It considers its own 13-tile hand and any open melds it has declared. It looks at every tile each opponent has discarded, and every meld each opponent has opened. It also tracks how many tiles remain in the wall — so it knows whether the game is early, mid, or nearly over.

Crucially, the AI only sees what you can see. Your concealed tiles are never revealed to it. It reads patterns from what is visible — just like a skilled human player reads the table.

Every Decision: Discard and Claim

Every action the AI takes at the table is driven by its model — there are no fallback scripted rules anywhere in the system.

Discard decisions make up the bulk of the AI's turns. Given its full game-state snapshot, the model produces a probability score for each possible tile to discard and plays the tile it judges most beneficial for its hand.

Claim decisions — whether to pung, chow, kong, or declare a win on another player's discard — account for roughly 30% of all in-game decisions. The same model evaluates each opportunity, weighing what it holds against what opponents have shown and how far the game has progressed. The result is an AI that defends intelligently, doesn't pung blindly, and times its wins the way an experienced player would.

The whole evaluation — for any decision — takes under a millisecond, so the AI never slows the game down.

No Two Games Feel the Same

A purely optimal AI would play the same way every time, which quickly feels mechanical and predictable. To avoid this, the AI uses a temperature setting that introduces natural variation into its choices.

Think of it this way: when the model evaluates a discard, it doesn't always pick its single top-ranked tile. Instead, it samples from a range of reasonable options — meaning it sometimes plays the second or third-best choice, just as a human player might. Higher temperature makes the AI more exploratory and unpredictable; lower temperature tightens its play and keeps it closer to optimal. Beginner games use a higher temperature so the AI feels natural and forgiving. Advanced games use a lower temperature for sharper, more consistent play.

Beyond individual AI variability, the opponents at a single table may operate at different difficulty levels — mixing strengths just like a real group of players. One opponent might play at Advanced level while another makes the occasional Beginner-style slip. This gives every game a more varied, human feel.

Two Difficulty Levels — Which Is Right for You?

Beginner

The Beginner AI learned from beginner-level games, so it plays at that level naturally — it builds sensible hands but makes the occasional realistic mistake. The temperature is set higher, making its decisions feel more human and forgiving. A good starting point if you are learning Hong Kong Mahjong or returning after a long break.

Advanced

Trained on high-skill gameplay with a 3+ faan minimum, the Advanced AI plays defensively, reads the table carefully, and wastes very few tiles. It is selective about claims, aware of opponent patterns, and times its wins the way an experienced player would. Its temperature is lower than Beginner, keeping its play tight and consistent. This is the opponent for players who want a genuinely tough game — and a useful benchmark for anyone who takes their mahjong seriously.

Play Offline — Same Hong Kong Mahjong AI

With Airplane Mode unlocked, the AI model runs entirely on your device. No server, no internet connection, no compromise. The opponents you face offline are the same AI — with the same model, the same difficulty levels, the same temperature variation — as the opponents in online games. Your next flight, commute, or off-grid weekend just became a mahjong session.

Frequently Asked Questions

Is the Hong Kong Mahjong AI actually intelligent, or does it just play random tiles?

The AI is genuinely intelligent. Every discard and every claim decision is driven by a machine learning model that reads the full game state — your visible discards, open melds, tiles remaining in the wall — and calculates the best move for its hand. It does not follow scripted rules or pick tiles at random.

Which difficulty should I choose?

If you are new to Hong Kong Mahjong or still learning the rules, start with Beginner. Once you can hold your own at a table, move to Advanced — a 3+ faan game with a sharp, defensive AI that reads the table carefully and rarely wastes a tile.

Can I play against the AI offline?

Yes. With Airplane Mode unlocked, the same AI model runs entirely on your device. No internet connection is required — the AI opponents are just as capable offline as they are in online games.

Does the AI cheat or know what tiles I have?

No. The AI only sees what any human player at the table would see: its own hand, all players' open melds, and the tiles that have been discarded. Your concealed tiles are never visible to the AI.

Ready to play against computer opponents that actually think?

Download the Hong Kong Mahjong app and start a game against the AI — Beginner or Advanced.

Related articles