Learning-Based Dynamic Allocation of Computational Resources for Cooperative Control Systems

Modern control architectures increasingly execute multiple control tasks on shared, and often remote, computational platforms, where a limited or expensive computing budget must be distributed among co-located applications. Conventional resource allocation strategies typically rely on explicit performance models and assume that individual systems can be evaluated independently. These assumptions become restrictive when several systems cooperate towards a common objective and their individual contributions cannot be measured in isolation. This thesis proposes a learning-based method for dynamic resource allocation in such cooperative settings. The method adopts a hierarchical architecture: a set of low-level control policies, one per system, drive the individual systems, while a high-level scheduling policy observes the global configuration of the environment and decides, at runtime, which systems to execute, thereby allocating the shared computational budget. Both levels are trained through reinforcement learning, allowing the scheduler to learn the mapping between the environment configuration and the criticality of each system without requiring an explicitly defined individual criticality metric. A single design parameter exposes the trade-off between resource utilization and control performance. The approach is developed incrementally across a series of environments and demonstrated on a cooperative multi-robot manipulation task, in which four robot arms jointly move an object to a target pose. The trained scheduler learns a configuration dependent activation strategy and, across the tested parameter settings, exhibits a monotonic resource-performance trade-off. This allows the designer to select between faster, resource intensive and slower, more resource-efficient operating points without re-modelling the systems or retraining the low-level control policies.