Massively Multiplayer Online (MMO) game servers must handle thousands of simultaneous players while maintaining sub-100ms response times. When server load exceeds capacity, traditional approaches either uniformly throttle all message types regardless of importance (damaging gameplay) or apply fixed heuristic rules that fail to adapt to dynamic workloads. This paper presents AFLL (Adaptive Feedback Loop Learning), a real-time load stabilization system that learns the causal relationship between outgoing server messages and subsequent incoming client requests. AFLL employs backpropagation to continuously adjust message type weights, enabling predictive throttling that blocks low-priority messages before overload occurs while guaranteeing critical message delivery. Through controlled experiments with 1,000 concurrent players, AFLL reduced average CPU time by 48.3% (13.2ms to 6.8ms), peak CPU time by 51.7% (54.0ms to 26.1ms), and thread contention by 64.4% (19.6% to 7.0%), while maintaining zero learning overhead through background computation and caching optimizations. The system achieved remarkable reproducibility (CV < 2% across all metrics) and identified a three-stage causal chain linking message blocking to load reduction. AFLL demonstrates that circular causality learning enables practical real-time adaptation for latency-critical systems.

翻译：大型多人在线（MMO）游戏服务器必须同时处理数千名玩家，并维持低于100毫秒的响应时间。当服务器负载超过容量时，传统方法要么不考虑重要性而统一限制所有消息类型（损害游戏体验），要么采用固定的启发式规则，无法适应动态工作负载。本文提出AFLL（自适应反馈循环学习），一种实时负载稳定系统，它学习服务器发出消息与后续客户端请求之间的因果关系。AFLL采用反向传播持续调整消息类型权重，实现预测性限流，在过载发生前阻断低优先级消息，同时保证关键消息的传递。通过1000名并发玩家的受控实验，AFLL将平均CPU时间降低了48.3%（从13.2毫秒降至6.8毫秒），峰值CPU时间降低了51.7%（从54.0毫秒降至26.1毫秒），线程争用降低了64.4%（从19.6%降至7.0%），同时通过后台计算和缓存优化实现了零学习开销。该系统实现了出色的可复现性（所有指标的变异系数CV < 2%），并识别出连接消息阻断与负载降低的三阶段因果链。AFLL证明，循环因果学习能够为延迟敏感系统实现实用的实时自适应。