PhD - Leuven | More than two weeks ago
Develop a hybrid ML and analytical mapping toolkit that unlocks state-of-the-art black-box optimization results.
Reinforcement learning (RL) algorithms have recently seen rapid advancement and adoption in a variety of fields, including but not limited to electronic design automation (EDA), mapping of tasks onto hardware resources, etc. Typically, the problem is represented as a graph (in many cases, directed acyclic graphs) and the goal is to transform or map this graph onto a limited set of resources. Each node on the graph might have certain properties which impose restrictions on which transformations it supports or on which resources can be mapped onto.
For these set of problems, traditionally heuristics have been used to find solutions that maximize or minimize some target criteria, such as data transferred, etc. Several successful machine learning approaches have been proposed for specific problems such as computer chip layout design, mapping tasks onto Machine learning accelerator, etc. However, these works primarily focus on a specific problem, even though many of these problems share similar underlying constraints. The primary challenge lies in representing the state-space of the system and how to efficiently encode the constraints so that valid solutions can be obtained.
In this project, the goal is to develop a hybrid RL and analytical mixed “mapping toolkit” that can be trained in reasonable time on public and proprietary benchmarks to unlock state-of-the-art black-box optimization results. The idea is to make the toolkit generic enough to allow different mapping problems to be solved using it with a straightforward user facing interface.
This project is an initiative of the Compute Systems Architecture Unit (CSA). CSA researches emerging workloads and their performance on large-scale supercomputer architectures for next-generation Artificial Intelligence (AI) and high-performance computing (HPC) applications. The team is responsible for algorithm research, runtime management innovations, performance modeling, architecture simulation and prototyping for these future applications and the future systems to execute them, to reach multiple orders of magnitude better performance, energy-efficiency, and total-cost-of-ownership.
Required background: Computer science, machine learning
Type of work: 60% modeling, 30% experimental,, 10% literature
Supervisor: Steven Latré
Co-supervisor: Debjyoti Bhattacharjee
Daily advisor: Debjyoti Bhattacharjee, Peter Vrancx
The reference code for this position is 2023-115. Mention this reference code on your application form.