Japan Advanced Institute of Science and Technology

Can We Infer Move Sequences in Go from Stone Arrangements?

The research investigates the feasibility of inferring the chronological move sequences of a Go game solely from the final arrangement of stones on the board.

Procedural Maze Generation with Considering Difficulty from Human Players' Perspectives

This research presents a novel method for procedural maze generation that incorporates human behavioral tendencies to calibrate difficulty levels.

• Researchers successfully calibrated procedural maze difficulty by training a LightGBM model on human navigation patterns, specifically focusing on path selection at branch points.
• The model was developed using data from 20 participants navigating 21 mazes, analyzing 23 distinct features such as distance to the goal and straight-line movement tendencies.
• Difficulty is quantified by the number of extra steps a player takes beyond the shortest possible path, allowing for the categorization of mazes into easy, moderate, and difficult tiers.

Improving Human Players' T-Spin Skills in Tetris with Procedural Problem Generation

This research explores the application of procedural content generation to create educational puzzles for Tetris, specifically targeting the mastery of the T-spin technique.

• Training with procedurally generated T-spin puzzles significantly increases maneuver frequency, with growth rates of 212% for random problems and 334% for curated 'interesting' problems compared to a decline in control groups.
• Procedural generation using a reverse-generation method effectively creates two-step T-spin puzzles that improve player performance in competitive Tetris scenarios.
• Machine learning models, specifically LightGBM, can predict puzzle difficulty and interestingness with high accuracy, achieving a symmetric mean absolute percentage error of 8.74% for interestingness ratings.

Design and Evaluation of a Cybersecurity Awareness Training Game

This study explores the design and evaluation of a serious game intended to improve cybersecurity awareness among university students.

• The serious game achieved an average score of 8.4 out of 10 for situational understanding and 7.6 for overall cybersecurity knowledge improvement among university students.
• The ATMSG (Activity Theory-based Model of Serious Games) framework successfully aligned educational objectives with game mechanics to teach complex security scenarios like phishing and public information safety.
• Beginner-level participants reported higher satisfaction with the game, finding the visual novel format both novel and accessible for learning cybersecurity concepts.

Deep df-pn and Its Efficient Implementations

Deep df-pn functions as an advanced iteration of the proof-number search algorithm, specifically engineered to resolve the seesaw effect that frequently degrades performance in game-tree searches.

• Deep df-pn resolves the 'seesaw effect' in game-tree searches by replacing static proof values with a depth-dependent function, preventing inefficient oscillation between subtrees.
• The algorithm utilizes two adjustable parameters to modulate search behavior, allowing for a more flexible balance between broad and deep exploration.
• Empirical testing in Connect6 demonstrates that Deep df-pn significantly reduces the number of nodes visited compared to standard df-pn and the 1+epsilon trick.

Machine-Learning of Shape Names for the Game of Go

This research addresses the challenge of enabling Go software to provide human-like commentary by automatically assigning appropriate shape names to specific moves.

• A machine-learning model using decision trees achieved an 82% accuracy rate in assigning human-like Go terminology to board moves.
• The model’s qualitative performance score of 83.8 is nearly identical to the 84.6 score achieved by human amateur labels when evaluated by professional players.
• The research utilized a dataset of 11,526 moves from 60 games, which were labeled by six players ranked 4 dan or higher using 71 predefined shape names.

Efficiency of Static Knowledge Bias in Monte-Carlo Tree Search

This analysis examines the integration of static knowledge into Monte-Carlo Tree Search (MCTS) algorithms for the game of Go.

• Integrating static move evaluation based on the Bradley-Terry model into MCTS algorithms significantly improves the competitive strength of Go engines against established programs like Fuego and Pachi.
• Progressive widening, which restricts search to a subset of promising moves, is highly effective for Go engines when tuned with a specific parameter, μ.
• The research validated these performance improvements through extensive testing of over 31,200 games on 13x13 boards, maintaining a 95% confidence interval of ±5% on winning rates.

Accelerated UCT and Its Application to Two-Player Games

The Accelerated UCT algorithm is introduced as a solution to the over-optimistic behavior and deceptive structures inherent in standard Monte-Carlo Tree Search (MCTS) when applied to two-player games.