TLDR: CogniPlay is a work-in-progress AI model designed to play games in a human-like manner, rather than just achieving superhuman performance. Inspired by the Double-Process Theory of Cognition, it combines intuitive, pattern-based decision-making (System 1) with selective, memory-bounded search (System 2). The model learns without human data, aiming for generalizability across various games, and will be evaluated using both quantitative and qualitative methods to assess its human-likeness.
While artificial intelligence has achieved remarkable feats in games like Chess, Go, and Dota 2, often surpassing human champions, these systems rarely mimic the way humans actually think and play. Traditional AI excels at brute-force calculations and vast memory, but human players rely on intuition, pattern recognition, and selective decision-making. This gap is what researchers Aloïs Rautureau and Éric Piette aim to bridge with their new work-in-progress model called CogniPlay.
The core idea behind CogniPlay is to create an AI that behaves more like a human in General Game Playing (GGP). GGP is a field focused on developing agents that can learn to play any game effectively, even with minimal prior knowledge. The current challenge with powerful AI opponents is that their difficulty often has to be artificially adjusted, for example, by limiting their search depth. A human-like AI, however, could offer more natural control over its strength and playstyle, making it more suitable for applications like replacing disconnected players in online games, enhancing non-player characters (NPCs), or even helping game designers automate playtesting.
CogniPlay draws inspiration from the Double-Process Theory of Cognition (DPTC), a psychological model that describes two distinct ways humans think: System 1 and System 2. System 1 is fast, intuitive, and relies on recognizing memorized patterns, much like how an experienced chess player might instantly recognize a common opening. System 2, on the other hand, is slower, analytical, and involves logical inference, used for deeper analysis in unfamiliar situations.
In CogniPlay, System 1 handles intuitive decisions by partitioning possible actions into ‘good’ and ‘bad’ subsets based on learned patterns. If System 1 is confident, it makes the move directly. If it’s uncertain, it triggers System 2. System 2 then performs a selective, memory-bounded Monte-Carlo Tree Search (MCTS), exploring only the ‘intuitively good’ actions identified by System 1. This aggressive pruning of the search tree is designed to replicate the selective search humans employ, who don’t exhaustively explore every possibility.
A key aspect of CogniPlay’s design is its ability to learn without needing human-generated data. System 1 learns from the decisions made by System 2 in novel situations, allowing the model to develop biases and playstyles that emerge algorithmically, rather than being explicitly programmed or imitated from human examples. This approach is crucial for its generalizability across a wide variety of games where human gameplay data might be scarce.
Measuring how ‘human-like’ an AI truly is presents its own set of challenges. The researchers plan to evaluate CogniPlay using both quantitative and qualitative methods. Quantitative tests will involve comparing the model’s decisions against observed human behavior, similar to how the Maia Chess engine was assessed. Qualitative tests, inspired by the Turing test, will involve human observers evaluating the AI’s playstyle without knowing if it’s human or machine, rating specific attributes like aggressiveness or tactical depth.
The initial implementation of CogniPlay is planned for the game of Renju, followed by integration into the Ludii game-playing system to test its generalizability across multiple games. The researchers acknowledge that developing generalizable opponent modeling methods is a promising future direction, as understanding an opponent’s thought process is a critical part of human problem-solving in games.
Also Read:
- Smarter AI Decisions: New Methods for Dynamic Abstraction in Monte Carlo Tree Search
- Bridging the Gap: A Hybrid AI Model for Understanding Human Minds
This research represents a significant step towards creating AI that not only plays games well but also plays them in a way that feels more natural and relatable to human players. For more details, you can refer to the full research paper: CogniPlay: a work-in-progress Human-like model for General Game Playing.
