Reproduce “A Unified Particle Filter LSTM for Data-Driven Process Simulation” on OpenResearch

About this paper

Parvin Malekzadeh, Opher Baron, Dmitry Krass

The Unified Particle Filter LSTM (Unified PF-LSTM) framework enhances data-driven process simulation by explicitly modeling latent-state uncertainty in event logs, improving the accuracy of simulated trajectories in complex operational settings like emergency departments. It achieves this by maintaining a weighted set of recurrent-state hypotheses, leading to reduced prediction errors for metrics such as Time to Provider Initial Assessment and Total Length of Stay.

Key insights

• Explicitly modeling latent-state uncertainty through a particle filter mechanism significantly improves the accuracy of data-driven process simulations, especially in environments with high partial observability.
• Leveraging MGF-based features to represent the particle belief provides a more comprehensive summary of the latent state distribution than a simple mean, allowing the model to capture nuanced aspects like dispersion or multimodality.
• A combined framework that handles routing and timing predictions, incorporates particle-based belief updates, and includes a robustness mechanism (routing validator) effectively learns complex, nonstationary, and history-dependent process dynamics directly from event logs.

Summary by alphaXiv.