Efficient Bayesian inference for stochastic agent-based models.

Area of Science:

• Computational modeling
• Bayesian inference
• Machine learning applications

Background:

• Real-world problems often require computationally intensive agent-based simulations.
• Parameter values in these models are frequently unknown and must be inferred from observed data.
• Traditional Bayesian inference relies on numerous simulations, leading to high computational costs.

Purpose of the Study:

• To compare machine learning methods for accelerating Bayesian inference in agent-based models.
• To investigate surrogate models and direct posterior estimation as alternatives to traditional simulation-based inference.
• To provide guidelines for selecting appropriate inference techniques for specific applications.

Main Methods:

• Development and application of surrogate models to replace computationally expensive simulations.
• Implementation of methods to directly estimate posterior distributions, bypassing traditional sampling schemes.
• Evaluation of machine learning approaches on stochastic simulations, including tumor growth and infectious disease spread models.

Main Results:

• Machine learning approaches achieve high inference accuracy with fewer simulations, offering significant speedups (orders of magnitude) over classical methods.
• No single algorithm universally outperforms others; method selection depends on the specific real-world application.
• The stochastic nature of phenomena can pose challenges for certain inference techniques.

Conclusions:

• Machine learning offers a computationally efficient alternative for parameter inference in complex agent-based models.
• Direct inference machines show particular promise for real-world applications.
• Practitioners should carefully consider the specific problem and its stochastic properties when choosing an inference method.