The paper introduces a novel framework for asynchronous Bayesian optimization within the Lava neuromorphic computing framework. Lava provides an abstract application programming interface for constructing event-based computational graphs, but existing solvers and optimization algorithms in Lava do not have the infrastructure to support event-based communication when problems are executed on separate compute nodes or architectures, leading to issues like deadlocking and wasted CPU cycles.
The proposed framework addresses these challenges by introducing an intermediate step between the optimizer and the black-box function. This step checks for stop or pause commands, handles the handshake operation to notify the main thread when the asynchronous search process is complete, and puts the process to sleep when the input port does not have any information, avoiding excess computation and deadlocks.
The authors showcase the capability of their asynchronous optimization framework by connecting Lava Bayesian Optimization (Lava BO) with a Quadratic Unconstrained Binary Optimization (QUBO) solver applied to a satellite scheduling problem, where the QUBO solver runs on Loihi 2 hardware. This test scenario highlights the ability of the proposed framework to support communication between multiple processes on different computing architectures where synchrony and runtime determinism are not guaranteed.
The authors plan to expand the framework by incorporating multiple agents communicating with a single optimizer and employing it to support lifelong, on-chip learning for robotics and signal processing applications.
Vers une autre langue
à partir du contenu source
arxiv.org
Questions plus approfondies