Continuous Reinforcement Learning for Queue-Based Service Scaling

The system implements self-service through autonomous reinforcement learning agents that automatically learn and execute scaling decisions without human intervention. The agents interact with the cloud environment, receive rewards based on scaling performance, and continuously improve their decision-making capabilities, enabling the system to adapt to unpredictable demand patterns while maintaining manageable complexity through automated self-optimization

The patent applies parameter changes by dynamically adjusting scaling parameters based on learned policies from reinforcement learning. The system modifies computing resource allocation parameters (such as number of instances, resource capacity) based on state observations and accumulated rewards, allowing adaptive response to demand fluctuations while the underlying learning framework manages the complexity of multiple parameter interactions

The system implements continuity of useful action through continuous reinforcement learning where the agent continuously accumulates reward values over time steps and continuously refines its scaling policies. This continuous learning process enables the system to progressively improve resource utilization efficiency while the ongoing nature of learning allows it to adapt to changing patterns without requiring complete retraining, thereby managing computational energy consumption

The patent applies preliminary action by pre-training the reinforcement learning model with accumulative reward mechanisms before deployment. The model undergoes preliminary training phases where it learns from simulated or historical data, accumulating reward signals to develop effective scaling policies. This preliminary action prepares the system for efficient operational decision-making, reducing the computational burden during actual production use