Simulation-based Optimisation using Differentiable Agents (SODA)

• Applicant: Dr. Philipp Andelfinger
• Subject Area: Data Management, Data-Intensive Systems, Computer Science Methods in Business Informatics
• Term: since 2022
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 497901036

Agent-based simulations are used in the design and evaluation of systems in various domains. Frequently, optimisation problems are solved by iteratively adjusting the parameters of the simulation model, which is associated with immense computation times. Gradient information, which could steer the optimisation process towards local optima, cannot be applied to most agent-based simulations as explicit gradient expressions are typically unavailable. The vision pursued in the project "Simulation-based Optimisation using Differentiable Agents" (SODA) is to rely on automatic differentiation techniques to enable gradient-based optimisation for a broad range of agent-based models. The key challenge is given by discrete model elements such as conditional branches, which are characteristic for agent-based models but can lead to unhelpful zero-valued or infinite partial derivatives. In a proof-of-concept publication, we showed that by manually substituting discrete model elements by smooth counterparts, the convergence speed and solution quality achieved in certain optimisation problems could be vastly improved without substantial deviations from the original agent behaviour. While this showed the significant promise of the approach on manually implemented example models, research is required to achieve a broad applicability and to systematically study the approach. Hence, the project addresses the following research questions: Can agent-based models be automatically translated to smoothed representations while retaining the per-agent behaviour of the original models? As a naive smoothing renders even small simulations computationally intractable, knowledge of common agent-based model elements must be exploited. Translation rules for model elements expressed in a domain-specific modelling language towards smooth simulation code will facilitate the automation. To what degree can the model properties and computational structure of agent-based simulations be exploited to reduce overheads? Unique opportunities for overhead reductions are given by properties such as sparsity in the agent interactions and structural similarities to artificial neural networks. Within a given time and compute budget, is the approach able to identify higher-quality solutions than existing methods? High-dimensional optimisation problems from transportation and cell biology serve as case studies to evaluate the approach.

• DFG Programme: Research Grants