Hard AI Improved (+1.8% Win Rate) & Audio Fix (v0.08b)

In this update, we have made the following two improvements:

• Adjusted the volume of the VGS-X Boot BIOS sound effects

• Improved Hard AI win rate: 56.5% → 58.3% (+1.8%)

VGS-X Boot BIOS Sound Adjustment

Previously, the startup sound of the VGS-X Boot BIOS (i.e., the virtual hardware layer) did not reflect the in-game volume settings.

With this update, the volume configured in the game is now directly applied to the hardware-level audio.

Hard AI Improvements

The Hard AI win rate has been improved from 56.5% to 58.3%.

Steam post imageAt first glance, a +1.8% increase may seem small, but for us, this represents a significant improvement.

The following algorithmic adjustments have been implemented:

• Refined card retention evaluation (slight win rate improvement)

• Strengthened stop-decision logic for high-probability 5-point hands (increasing expected value)

• Added incentives to interfere with opponent’s fixed yaku (5-point hands) (reducing opponent expected value)

• Introduced a delay strategy to increase the probability of reaching 7 points when a 5-point hand is already secured (increasing expected value)

• Added incentives to continue Koi-Koi even for marginal +1 point gains when risk is extremely low (increasing expected value)

• Improved counterplay evaluation during opponent’s Koi-Koi by considering both captured and played cards (reducing opponent expected value)

• Introduced bluff strategies that intentionally allow the opponent to score small points (reducing opponent expected value)

• Improved early detection of danger when the opponent has 2/3 cards for a 5-point yaku and the remaining card is still in play (reducing opponent expected value)

• Added mild suppression against opponent’s 6-point Koi-Koi extensions (reducing opponent expected value)

• Added logic to stop low-growth Koi-Koi situations (reducing risk from x2 multipliers)

• Penalized excessive Koi-Koi in late-game situations (favoring closing out wins)

• Added incentives to continue Koi-Koi when growth potential is high, even for small gains (increasing expected value)

Notes

In games such as Mahjong, Poker, and Hanafuda, when two players of perfectly equal skill compete, the win rate converges to 50%.

Within that 50%, a certain portion of outcomes is effectively determined by initial conditions such as the dealt hand or draws.

I refer to this portion as the “天賦率" (Luck Ratio.)

For example, if a game has a Luck Ratio of 40%, then the remaining 20% represents the portion of the outcome that can be influenced by strategy (skill).

In other words, in a game with a 40% Luck Ratio, the theoretical maximum win rate is 60%.

In human-versus-human matches, psychological factors also come into play, making it difficult to measure this value precisely. However, in AI-versus-AI matches, there is no psychological noise. By developing AI that does not cheat (i.e., does not access hidden information such as the deck or the opponent’s hand), it becomes possible to estimate the game’s intrinsic Luck Ratio.

This AI improvement is part of an ongoing effort to explore the Luck Ratio in Hanafuda (Koi-Koi).

To the best of my knowledge, no one has yet identified this value for Hanafuda.

If you are aware of any prior work on this topic, I would greatly appreciate hearing about it.

While it is theoretically possible to calculate an exact Luck Ratio using statistical methods, exhaustively simulating all possible deal patterns over 12-round matches would likely require an impractical amount of time, even with modern supercomputers.