Active Inference for Dynamic Bayesian Networks

Caner KomurluIJCAI 2016

Abstract

This paper explores the application of active inference to Dynamic Bayesian Networks (DBNs) to optimize observation selection in resource-constrained environments. By strategically choosing which variables to observe at runtime, the proposed methods significantly reduce prediction uncertainty and resource consumption. The approach is validated through simulations in tissue engineering and real-world sensor network data, demonstrating that network-aware selection strategies outperform traditional baselines as observation budgets increase.

Introduction

In many real-world systems, continuous and comprehensive observation is often infeasible due to physical or economic constraints. For instance, in wireless sensor networks (WSNs), constant data transmission rapidly depletes battery life, while in tissue engineering, monitoring biological growth is both costly and technically complex.

While standard machine learning typically focuses on active learning during the training phase, this research addresses active inference during the prediction phase. The goal is to determine, given a limited budget for observations, which specific variables should be monitored to best predict the state of the entire system. While DBNs are powerful tools for modeling temporal changes, they suffer from increasing uncertainty over time without new evidence. This work introduces strategies to pick the most "informative" observations to maintain model accuracy.

Methodology

The research applies active inference to DBNs across two distinct domains:

1. Tissue Engineering (Vascularization)

The system models blood vessel growth on a grid, where the DBN captures how growth in one cell depends on its neighbors from the previous time step. The objective is to identify the optimal time ($t^*$) to perform an observation to ensure that final prediction uncertainty remains below a critical threshold. This allows researchers to pinpoint the exact moment an observation provides the most value for ensuring a successful biological outcome.

Uncertainty reduction over time for different biological stress levels in tissue engineering simulations.

2. Wireless Sensor Networks (WSNs)

In WSNs, the system observes only a subset of sensors (defined by a budget $B$) to predict the state of the rest of the network. The author introduces Net Impact-Based Selection (NBS), a novel method that prioritizes variables based on their "net impact" on neighbors within the DBN structure. This method leverages network topology, recognizing that nodes with strong dependencies provide more information about the overall system than isolated nodes.

Key Results and Insights

WSN Performance and Scaling

The effectiveness of the NBS method is particularly evident as the observation budget increases. The study compared NBS against several baselines, including Random Selection (RND), Sliding Window (SW), Gaussian Processes (GP), and Kalman Filters (KF). While traditional methods like GP and KF perform well under very tight budgets (e.g., 10%) due to their proficiency with local attributes, the DBN-based NBS method scales superiorly. As the budget grows from 20% to 50%, NBS consistently outperforms or matches all baseline methods in reducing prediction error.

Comparison of prediction error across different observation budgets for temperature data.

Uncertainty Management

In the tissue engineering simulations, the model successfully mapped non-linear uncertainty curves for different biological stress levels (Low, Moderate, High). By identifying that uncertainty reduction is not linear, the system can optimize the timing of data collection to maximize resource efficiency.

Conclusion and Future Work

The research demonstrates that shifting focus from model building to smart real-time data collection can significantly improve system monitoring. However, computing uncertainty for every possible time slice is computationally expensive. Future work will focus on developing more efficient search methods and moving toward "information fusion," where a single DBN handles multiple data domains (e.g., temperature and humidity) to exploit cross-domain correlations. Additionally, the author plans to explore advanced selection criteria such as Maximum Variance Selection and Expected Variance Reduction.

Summary updated

BibTeX
@inproceedings{komurlu2016active,
  author = {Caner Komurlu},
  title = {{Active Inference for Dynamic Bayesian Networks}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2016},
}
Figure