박준영 교수

We propose a framework to learn to schedule a job-shop problem (JSSP) using a graph neural network (GNN) and reinforcement learning (RL). We formulate the scheduling process of JSSP as a sequential decision-making problem with graph representation of the state to consider the structure of JSSP. In solving the formulated problem, the proposed framework employs a GNN to learn that node features that embed the spatial structure of the JSSP represented as a graph (representation learning) and derive

We propose ScheduleNet, a RL-based real-time scheduler, that can solve various types of multi-agent scheduling problems. We formulate these problems as a semi-MDP with episodic reward (makespan) and learn ScheduleNet, a decentralized decision-making policy that can effectively coordinate multiple agents to complete tasks. The decision making procedure of ScheduleNet includes: (1) representing the state of a scheduling problem with the agent-task graph, (2) extracting node embeddings for agent an

A low-power real-time traffic sign recognition system that is robust under various illumination conditions is proposed. It is composed of a Retinex preprocessor and an SVM processor. The Retinex preprocessor performs the Multi-Scale Retinex (MSR) algorithm for robust light and dark adaptation under harsh illumination environments. In the Retinex preprocessor, the recursive Gaussian engine (RGE) and reflectance engine (RE) exploit parallelism of the MSR tasks with a two-stage pipeline, and a mixe

WebAssembly runtimes embed compilers to compile WebAssembly code into machine code for execution. These compilers use various compiler rules to define how to optimize and lower the WebAssembly code. However, existing testing tools struggle to explore these rules effectively due to their complexity. Moreover, they cannot generate test cases diversely due to their limitations, which can result in undetected bugs. This paper presents RGFuzz, a differential fuzzer for WebAssembly runtimes, addressin

Multiple granularity parallel core architecture is proposed to accelerate object recognition with low area and energy consumption. By adopting task-level optimized cores with different parallelism and complexity, the proposed processor achieves real-time object recognition with 271.4 GOPS peak performance. In addition, content-aware fine-grained task scheduling is proposed to enable low power real-time object recognition on 30fps 720p HD video streams. As a result, the object recognition process

Combinatorial Optimization (CO) problems are theoretically challenging yet crucial in practice. Numerous works used Reinforcement Learning (RL) to tackle these CO problems. As current approaches mainly focus on single-worker CO problems such as the famous Travelling Salesman Problem (TSP), we focus on more practical extension of TSP to multi-worker (salesmen) setting, specifically MinMax mTSP. From the RL perspective, Minmax mTSP raises several significant challenges, such as the cooperation of

Heterogeneous multi-core object recognition processor with Reinforcement Learning (RL) NoC is proposed for efficient portable HD object recognition. RL NoC automatically learns management policies in the network of heterogeneous system without an explicit modeling. By adopting RL NoC, the throughput performances of feature detection and description are increased by 20.4% and 11.5%, respectively. As a result, the overall execution time of the object recognition is reduced by 38%. The implemented

We propose ScheduleNet, a scalable scheduler that minimizes task completion time by coordinating multiple agents. We formulate the min-max Multiple Traveling Salesmen Problem (mTSP) as a Markov decision process with an episodic reward and derive a scalable decision-making policy using Reinforcement Learning (RL). The decision-making procedure of ScheduleNet includes (1) representing the state of a problem with the agent-task graph, (2) extracting node embedding for agents and tasks by employing

With the growing importance of sustainable development, ESG (Environmental, Social, and Governance) management has become a critical component of corporate strategy. Despite this, the domestic ESG evaluation system encounters significant challenges due to inconsistencies in evaluation criteria and the heterogeneous use of data among ESG rating agencies, which compromise the credibility of the evaluations. This study addresses these issues by proposing a methodology that utilizes Large Language M

This paper presents a new technique for generating precise clock delays. The technique can obtain finer timing resolution less than the gate delay of the delay chain by locking in multiple clock period. Using this technique, a 250ps of timing resolution could be achieved from a 750ps delay of the single delay stage in a DLL(Delay Locked Loop) structure. The delay chain of the proposed circuit is locked on three times of the clock period and a finer delay resolution than the absolute gate delay i

This paper suggests Deep Neural Controller (DNC), a network architecture for sequential decision making based on high order Markovian state-space model. DNC is composed of two components, one for modeling system dynamics and another for constructing decision making policy. In this architecture, deriving control policy is conducted by training DNC network. We first employ a deep neural network to model the dynamic behavior of a complex dynamic system that has high-order Markovian property. By int

탈중앙 연합학습은 중앙서버 없이 각 노드 간의 통신을 통해 모델을 교환하고 결합하는 분산 협력 학습방식이다. 기존의 연합학습과 비교했을 때, 단일 장애점, 병목 현상 문제를 극복할 수 있다는 장점이 있다. 하지만 탈중앙 연합학습은 어그리게이터의 부재로 인한 학습성능 저하를 겪게 된다. 또한, 연합학습과 같은 분산학습에서는 모델의 일반화 역량 및 개인화 역량 모두 갖춘 모델을 학습하는 것이 중요하다. 본 연구는 어그리게이터 노드를 도입하여 어그리게이터의 부재로 인한 학습 성능 저하를 극복하고, 개인화 역량 및 일반화 역량을 모두 갖출 수 있는 클러스터링 기반 탈중앙 연합 학습 프레임워크를 제안한다. 본 논문에서 제시한 방법은 어그리게이터를 효과적으로 도입하기 위해 학습 방향 유사도를 기반으로 한 클러스터링을 진행한다. 클러스터 내 학습에서는 어그리게이터 노드를 활용하여 유사한 학습 방향을 가진 노드들과의 협업을 통해 개인화 역량을 갖추고자 한다. 이후 지나친 개인화 방지 및 일반화 역