Discrete-Event-Simulation Based on Machine Learning Predictive Agents

The use of simulators for the generation of synthetic data is an important piece to validate data processing systems in real time, such as Online Failure Predictors. The prediction models used in these systems are capable of recognizing in their activity patterns, the necessary clues to anticipate the appearance of events. In this work, we present an agent-based simulator combining machine learning models and statistical models for different types of events. The key contribution of our model is the definition of a feedback loop between the agents output and input, enabling the synthetic generation of rich datasets that maintain the learned relationships from the original data.