Adaptive Pattern Reduction for Long-Term Speculative Cloud Gaming Systems

Speculative execution in cloud gaming pre-renders predicted frames to mitigate network latency, but faces exponential computational costs and static prediction models that cannot adapt to player behavior. This paper proposes an adaptive retraining framework for LSTM-based pattern prediction that dynamically updates models using real-time user input. Experiments across three game genres demonstrate efficient pattern reduction, achieving performance comparable to that of pre-trained models with short training intervals.