Hard AI Fix in Watch Mode + Performance Improvement (v0.10a)

In this update, we’ve fixed an issue with the Hard AI in Watch Mode and made a small improvement to its win rate.

• Fixed an issue where the Hard AI did not correctly track the number of cards remaining in hand

• Recording the Random Seed in watch.log

• Improved Hard AI win rate vs Normal: 57.6% → 57.7% (+0.1%)

■ Hard AI Bug Fix

We identified and fixed an issue in Watch Mode where the Hard AI could behave inconsistently with its intended logic, sometimes resulting in poor decisions.

WHAT WAS HAPPENING

In Watch Mode, the AI incorrectly treated its hand as if it always had 8 cards (the initial state).

The Hard AI evaluates which yaku are still achievable and estimates their expected value (based on how many turns it may take to complete them), using the number of cards remaining in hand.

Due to this bug, the AI could mistakenly assume that even very unlikely yaku were still achievable, which prevented it from performing at its intended strength.

This issue has now been resolved:

https://github.com/suzukiplan/open-hanafuda-ai/commit/2ac32ae16f8592b0568998f5ca3525b2dc37f606

■ Recording the Random Seed in watch.log

We have implemented a fix to record the random seed value in `watch.log`, which is generated when observing AI vs AI matches in Watch Mode.

The development team plans to use `watch.log` files provided by users via GitHub Issues and other channels to analyze strategic weaknesses and improve the Open Hanafuda AI algorithms.

In this type of AI improvement, it is crucial to run reproducible tests to verify how changes actually affect match outcomes. To support this, the Battle Hanafuda development team has developed a replay tool that can reconstruct identical game states from the data recorded in `watch.log`.

To accurately reproduce matches using this replay tool, the random seed value is required.

This replay tool has also been released as part of Open Hanafuda AI.

■ Hard AI Win Rate Improvement

We also made a small adjustment that improved the Hard AI’s win rate against Normal from 57.6% to 57.7%.

Details:

https://github.com/suzukiplan/open-hanafuda-ai/issues/2

■ About Open Hanafuda AI

The AI system used in Battle Hanafuda, “Open Hanafuda AI,” is available as open source under the MIT License.

It is based on Monte Carlo methods rather than Deep Learning, making it a good reference for those interested in traditional AI techniques.

Feel free to check it out, and contributions (issues / pull requests) are always welcome:

https://github.com/suzukiplan/open-hanafuda-ai