Technology Category
Machine Learning: Evolutionary LearningMultiagent Systems: Adversarial AgentsGame Theory and Economic Paradigms: Adversarial Learning
Evolutionary Optimization of Deep Learning Agents for Sparrow Mahjong
🤖 AI Summary
This work addresses the challenging domain of Sparrow Mahjong—a stochastic, partially observable multi-agent game. We propose Evo-Sparrow, an end-to-end gradient-free AI agent. Methodologically, it employs an LSTM to model sequential game states and directly optimizes policy parameters via the Covariance Matrix Adaptation Evolution Strategy (CMA-ES), bypassing reliance on environment modeling or gradient estimation inherent in conventional reinforcement learning. Crucially, the framework requires neither reward shaping nor value-function approximation, enhancing both training stability and interpretability. Empirical evaluation demonstrates that Evo-Sparrow significantly outperforms random and rule-based baselines in large-scale simulations, matches PPO’s performance, and exhibits more robust convergence. To our knowledge, this is the first evolutionary optimization–based pure sequential decision-making paradigm for imperfect-information stochastic games, extending the applicability of deep learning to complex strategic reasoning under uncertainty.
Technology Category
Application Category
📝 Abstract
We present Evo-Sparrow, a deep learning-based agent for AI decision-making in Sparrow Mahjong, trained by optimizing Long Short-Term Memory (LSTM) networks using Covariance Matrix Adaptation Evolution Strategy (CMA-ES). Our model evaluates board states and optimizes decision policies in a non-deterministic, partially observable game environment. Empirical analysis conducted over a significant number of simulations demonstrates that our model outperforms both random and rule-based agents, and achieves performance comparable to a Proximal Policy Optimization (PPO) baseline, indicating strong strategic play and robust policy quality. By combining deep learning with evolutionary optimization, our approach provides a computationally effective alternative to traditional reinforcement learning and gradient-based optimization methods. This research contributes to the broader field of AI game playing, demonstrating the viability of hybrid learning strategies for complex stochastic games. These findings also offer potential applications in adaptive decision-making and strategic AI development beyond Sparrow Mahjong.
Problem
Research questions and friction points this paper is trying to address.
Optimizing AI decision-making in Sparrow Mahjong using deep learning
Evaluating board states in non-deterministic, partially observable environments
Combining evolutionary optimization with deep learning for strategic AI
Innovation
Methods, ideas, or system contributions that make the work stand out.
LSTM networks optimized by CMA-ES
Deep learning with evolutionary optimization
Hybrid strategy for stochastic games
🔎 Similar Papers
No similar papers found.
