In this update, we have added two new modes: Hard Mode and No-Sake Mode.
Hard Mode: Play against a Hard AI designed to achieve “human-level strength” without any form of cheating
No-Sake Mode: Play under a ruleset without the Sake Cup
Originally, this update was scheduled for mid-April, but it has been slightly moved forward.
The reason for this change is explained in the No-Sake Mode section below.
This new mode allows you to directly play against the Hard AI, which was previously only available for observation in Watch Mode.
In many Hanafuda video games, difficulty is often adjusted using methods that are effectively impossible for human players, such as:
Favorable initial hands
Seeing the opponent’s hand
Knowing future draws
In other words, forms of “cheating.”
However, the Hard AI in Battle Hanafuda does not rely on any of these.
Instead, it is built on:
Monte Carlo simulation-based decision making (expected value estimation)
Heuristic strategies derived from understanding the core structure of Koi-Koi as a game
By combining these, the AI is designed around a single principle:
A match that is fully driven by reasoning under imperfect information.
To ensure transparency (i.e., to prove that no cheating is involved),
the entire AI logic is published as open source:
https://github.com/suzukiplan/open-hanafuda-ai
At this level of implementation, examples in Hanafuda games are extremely rare.
At the very least, I have not been able to find other implementations following a similar approach.
While such cases do exist in Mahjong software,
Hanafuda is a more niche domain, and it is likely difficult to justify this level of R&D investment in commercial projects.
In my case, however, this is a personal project.
The cost of developing this AI is roughly $20/month for Codex.
(Codex is an excellent engineer!)
Currently, many players who try other Hanafuda games tend to think:
“Hanafuda is just a game of luck”
“Video game AIs are all cheating anyway”
However, Battle Hanafuda is one of the few games that strictly follows the modern Nintendo rules, which are designed in a way that allows decision quality to meaningfully influence outcomes, even under inherent randomness.
The Hard AI is continuously being optimized specifically for these rules.
If you find that this is actually the case through your own play, I would appreciate it if you share your thoughts via discussions or reviews.
Most developers (myself included) tend to believe that their own game is the best.
Feedback from third parties is the most effective way to correct that bias.
This mode allows you to play with a 46-card deck, excluding:
The September (Chrysanthemum) Sake Cup
One additional Kasu card from September
Many Hanafuda games include this option, so it was added to better accommodate player expectations.
However, I previously stated in a Steam News post that:
“No-Sake Mode increases skill dependency”
After further analysis, I found that the opposite is true:
No-Sake Mode actually reduces skill dependency.
In a 12-round game, each player has roughly two opportunities on average to receive the Sake Cup at the start of a round.
The Sake Cup increases the number of strategic branches,
which allows skill differences to more strongly affect outcomes.
With Sake Cup: Hard AI win rate = 60.8%
Without Sake Cup: Hard AI win rate = 56.0%
※ In modes that are not multiples of 6 rounds, the distribution of initial hands becomes biased.
(Under Nintendo rules, 12 rounds is the standard format.)
A more detailed analysis is available here (Japanese):
https://note.com/suzukiplan/n/n6d296ea19b15
The development team prioritizes adherence to the modern official (Nintendo) rules.
As a result, we decided to cancel further AI improvements specific to No-Sake Mode,
and instead focus development resources on improving the Standard Mode AI.
This allowed us to move the release slightly earlier than planned.
With the addition of new modes, both leaderboards and achievements have been expanded.
Leaderboards: +12 (4 → 16)
Achievements: +12 (28 → 40)
Each mode now includes variants for 2 / 4 / 6 / 12 rounds.
The win rate of Hard AI against Normal AI has improved to 62.1%.
This is an increase of +1.8 pts (~+3%) from version 0.11 (60.3%).
Details can be found here:
https://github.com/suzukiplan/open-hanafuda-ai/milestone/3?closed=1
Further improvements to Hard AI are ongoing.
This might be a good opportunity to unlock achievements more easily… for now.
The method for exiting the game has been changed:
Previously: Select “Exit” from the title menu
Now: Press B on the “Press Any Button” screen, then press A
Previously: 4 rounds / Normal / Standard
Now: The last played Free Play mode
Previously: Loaded on startup or after score submission
Now: Loaded when opening the leaderboard screen or submitting a score
This change reduces load on Steam servers due to the increase from 4 → 16 leaderboards.
Before developing Battle Hanafuda, I did not fully understand Koi-Koi myself.
The Normal AI represents how I used to think as a beginner.
The Hard AI represents how I think now.
If you find the Hard AI still too weak, that likely means your understanding of Hanafuda has already surpassed mine.
I encourage you to try Hard Mode and see for yourself.
Finally, a brief note.
The Open Hanafuda AI used in Battle Hanafuda is provided as free software under the MIT License.
This means it can be freely used not only in Battle Hanafuda, but also in other Hanafuda video games.
If the goal is to eliminate the “misconceptions” and “resignation” surrounding Hanafuda video games mentioned earlier, I believe it may be more effective for this algorithm to be adopted in titles with greater commercial reach than my own.
Of course, contributions such as improvement proposals via Issues or Pull Requests to Open Hanafuda AI are welcome.