Back to search

Democritus University of Thrace (DUTH)

Integrated circuits design automation with reinforcement learning

Abstract

dc:description

The increasing complexity and scale of modern integrated circuits demand new methodologies that can optimize both performance and design effort in Electronic Design Automation (EDA). Traditional approaches in physical synthesis often rely on hand-crafted heuristics and fixed optimization flows, which can be suboptimal and difficult to generalize across design contexts. Recent advances in machine learning and particularly in Reinforcement Learning (RL), offer a powerful alternative. RL enables adaptive decision-making through continuous learning from feedback, making it an ideal candidate for navigating the vast and discrete optimization spaces characteristic of EDA problems. This thesis explores the integration of RL into two critical areas of digital design: timing closure in physical design and approximate datapath synthesis.The first part of this thesis focuses on timing closure, a core challenge in the physical design flow. We propose a novel optimization methodology that combines a Lagrangian Relaxation (LR)-based cost modeling framework with a Multi-Armed Bandit (MAB) algorithm to guide the selection of timing optimization heuristics. This hybrid approach enables the autonomous execution of a rich set of optimization techniques, including gate sizing, flip-flop resizing, buffer insertion, pin swapping, gate merging/splitting, and the application of useful clock skew. Each heuristic is dynamically selected in a context-aware manner, based on its historical reward in terms of timing improvement versus computational cost. The result is a fully adaptive and self-tuning optimization loop that eliminates the need for manual flow tuning.The second part of the thesis investigates the use of RL for the design of application-specific approximate arithmetic units, with a focus on parallel prefix adders. Approximate computing is increasingly relevant for domains where energy efficiency is prioritized over full numerical accuracy. We introduce an RL framework that co-optimizes hardware efficiency and application-level performance. An RL agent learns to explore the design space of approximate adders by receiving feedback from both hardware synthesis metrics and application error behavior. In contrast to previous approaches that rely on predefined architectural templates, our method employs a novel enumerative synthesis technique that generates all structurally valid solutions satisfying given constraints. This allows the discovery of Pareto-optimal designs with diverse trade-offs in delay, area, power, and accuracy across a variety of real-world applications.In summary, this thesis demonstrates the power of Reinforcement Learning and statistical decision-making to redefine traditional EDA methodologies. By intelligently orchestrating optimization decisions in both physical design and datapath synthesis, the proposed methods achieve state-of-the-art results while significantly reducing manual design effort. These contributions pave the way for more autonomous, efficient, and intelligent design tools in future VLSI systems.

Degree

thesis:*
Grantor dc:publisher
Democritus University of Thrace (DUTH)
Year dc:date
2025

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Stefanidis, Apostolos
  • Στεφανίδης, Απόστολος

Subjects

dc:subject × 20

Rights

Language dc:language
eng

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/59805

Chain of custody

source
Harvested from
Greek National Archive of PhD Theses
Base URL
phdtheses.ekt.gr/eadd_oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
citation

Stefanidis, Apostolos; Στεφανίδης, Απόστολος. Integrated circuits design automation with reinforcement learning. Democritus University of Thrace (DUTH), 2025. http://hdl.handle.net/10442/hedi/59805