Project summary

Ordinary differential equation (ODE) models are essential in many research areas, such as systems biomedicine, to investigate and understand dynamical systems. For parameter optimization, often only a small amount of experimental data is available, with data on varying biological conditions that provide independent information being particularly scarce. Although we could identify several pitfalls and could suggest several methods, the availability of reliable methods for numerical parameter optimization is still a major bottleneck. To improve the reliability of parameter optimization, we will develop a meta-learning approach that learns to solve an optimization task at hand from preceding optimization steps and similar other optimization problems. Specifically, we will apply reinforcement learning for a problem-specific selection of an appropriate optimization method and for tuning of the hyperparameters for the optimization and ODE integration algorithms. Publicly available benchmark problems with experimental and simulated data will be used to derive policies in the form of actions that choose the best optimization method.

Our methods

• Combining knowledge- and data-driven modeling
• Differential equations
• Meta-learning
• Pre-training
• Local perspective