New AI Model Mimics Human Moves in Xiangqi

Xiangqi, also known as Chinese Chess, is a popular board game in China. It has a long history and involves players simulating a battle to checkmate their opponent’s king. The game offers many strategies and can be quite complex. With advancements in artificial intelligence (AI), researchers have been developing ways to create computer programs that can play board games better than humans. Most of these programs focus on winning, but there is also a growing interest in making AIs that play more like humans.

This article discusses a new AI model called Xiangqi Structurally Variable (XQSV). This model aims to imitate how human players play Xiangqi by predicting their Moves based on their behavior. The goal of XQSV is not just to win games but to do so in a way that feels Human-like.

#Board Game Engines

AI has made significant progress in board games. Programs like DeepMind's AlphaGo have shown that AI can outperform human experts. These programs learn the game rules and practice through self-play, improving their skills over time. Other engines, like chess engines TDLEAF, focus on different learning strategies, but the goal remains the same: to become better than human players.

However, many programs prioritize winning and do not focus on how humans play. For example, traditional engines may struggle to predict human moves because they follow strict rules, while humans can be more flexible in their thinking. This provides an opportunity to build AI that mimics human behaviors instead of simply optimizing winning strategies.

#The Need for Human-Like AI

Creating an AI that plays like a human has several benefits. Such AIs can provide more enjoyable experiences for players. They can also serve as coaches, giving strategic advice and helping players improve their skills. Additionally, a human-like engine can be integrated into existing AI systems to improve their evaluations of moves.

Most current methods for making AI behave like humans involve analyzing past games and predicting potential moves. One notable program, Maia, predicts moves based on human chess games and has achieved better Accuracy than some traditional engines. However, the pursuit of a human-like game AI in Xiangqi has not been extensively studied.

#Xiangqi and its Complexity

Xiangqi has a rich history and offers a challenging environment for AI research. The game involves complicated strategies and numerous potential moves, making it an excellent platform for studying decision-making processes. Most existing Xiangqi engines use self-play and game tree search techniques to learn but do not focus on mimicking human behaviors.

The new XQSV model fills this gap. It aims to set a standard for future research on AI that can replicate the way humans play Xiangqi.

#XQSV - The Model

XQSV adopts a unique approach to predict human moves in Xiangqi. The model uses a specific structure that can change based on the skill levels of players it encounters. This allows it to adapt to different player abilities, which is critical since advanced players often think differently than beginners.

#Data Preparation

To begin with, XQSV retrieves Xiangqi game data in a specific format. The data is then preprocessed by breaking it into parts. This involves organizing moves based on the skill levels of players, which are ranked using the Elo rating system. By partitioning data based on skill levels, the model can better learn and predict player behavior.

#Sequential Input

Many current game AIs rely on a two-dimensional board layout to remember past moves. However, in interviews with actual Xiangqi players, it was found that most of them recall previous moves in a linear order. Thus, XQSV takes input as a sequence of moves instead of board configurations. This decision aligns the model more closely with the way humans think while playing.

#Memory Limitations

XQSV works under the assumption that human players do not remember every move from the start to the end of a game. Therefore, the model limits its memory to a certain number of past moves. This design choice also enhances computational efficiency and mimics the natural limitations of human memory.

#Dynamic Network Structure

One innovative feature of XQSV is its ability to adjust its network structure based on the data it processes. Different players have different styles, and by changing its architecture depending on the skill level of the player, XQSV can more accurately imitate human behavior.

#Filter for Illegal Moves

A major challenge in predicting moves is the vast number of possible actions. XQSV reduces this by introducing a filter that only allows legal moves based on the current board state. This helps to narrow down the choices and improves the accuracy of the predictions.

#Experimentation and Results

To evaluate XQSV, tests were conducted on different skill levels. The experiments confirmed that as player skill increases, a more complex model structure is required to simulate their gameplay. The results indicated that XQSV achieved about 40% accuracy overall when predicting human moves.

#Turing Test

A significant part of the evaluation involved a Turing Test, where real human players were asked to differentiate between human opponents and the XQSV AI. Participants played games against three opponents: a human player, the XQSV model, and a standard mobile app. Results showed that players had trouble distinguishing between the real player and XQSV, highlighting the model's effectiveness in imitating human behavior.

#Non-Deterministic Nature of Human Play

One challenge in creating an AI that plays like a human is the unpredictable nature of human decisions. The model can become deterministic, only predicting the best move without considering variability that multiple human players might exhibit. To address this, XQSV generates a probability distribution for potential moves, making it possible to introduce randomness to mimic human-like play more closely.

#Conclusion

XQSV represents a step forward in developing AI that imitates the way humans play Xiangqi. Through various design choices, including filtering illegal moves, partitioning data, and adjusting memory limits, the model shows promise in predicting moves similar to a human player. XQSV achieves a notable accuracy of around 40%, which can increase when considering relaxed evaluation metrics.

In summary, this model could have practical applications in enhancing the experience of playing Xiangqi and could serve as a foundation for future research into the development of human-like AIs in board games. As the field progresses, further improvements could lead to even more sophisticated models capable of mimicking the complexities of human thought processes in gameplay.