56016 research outputs found
Sort by
Empirical Study on Contextual Bandit Algorithms with Neural Processes
No full textGraduate School of Artificial IntelligenceMulti-armed bandit is a well-formulated test bed for designing sequential decision-making algorithms that deal with the exploration-exploitation dilemma. Bandit algorithm balances exploration towards uncertain domain and exploitation of the observed history to accurately estimate the reward distribution of each arm. Contextual bandit incorporates a context that contains rich information about the structure of bandit environment and determines the reward function, making the algorithms devised in this setting be more applicable to real-world problems like news recommendation. Focus on the practicality of contextual bandit algorithm, we consider the diversity and non-stationarity of bandit environment. Also, we assume that there exists a number of accumulated dataset from previous evaluations. To this end, we propose offline training of a reward prediction model via meta-learning so that the model can adapt to the changing environment. We consider Neural Processes (NP), a probabilistic few-shot learner that can estimate the uncertainty with its prediction. Adopting the upper confidence bound (UCB) exploration strategy, we propose NP-UCB, the exploration strategy based on the uncertainty estimate of trained neural processes. We evaluate the proposed algorithm with various neural processes on wheel bandit and news recommendation system. The results show that our method works well with the latest neural process model called Neural Bootstrapping Attentive Neural Processes (NEUBANP), which can adapt to dynamically changing environments with the help of its reliable uncertainty estimates.ope
Factors Influencing the Public Acceptance of the Floating Offshore Wind Farm Development in Ulsan
Get PDFGraduate School of Technology and Innovation ManageThis thesis investigates the determinants of the public acceptance of the floating offshore wind farm development in Ulsan, highlighting the role of the community profit-sharing schemes. The Ulsan project is currently the largest floating offshore wind farm plan in the world and aims to install a total capacity of 9GW by 2030. As many cases at home and abroad demonstrated, public acceptance is a key success factor of large-scale renewable projects. Although there are many factors influencing public acceptance, profit-sharing schemes for local communities have recently drawn a significant attention from local residents and policy-makers alike. This thesis examines the individual effects of profit-sharing schemes and further their joint effects with other factors such as prior knowledge, trust, distributive and procedural justice, perceived risks and perceived costs, and perceived benefits. We collected data from a survey of 500 individuals who live in Ulsan and analyzed the data using OLS. The theoretical and policy implications of our empirical findings are discussed.ope
??????????????? ????????? ??????????????? ????????? ????????? ?????? ???????????? ????????????
No full textDepartment of Materials Science and Engineeringclos
Human ATAD5-RLC forms two distinct complexes functioning in PCNA unloading and Ub-PCNA de-ubiquitination.
No full textDepartment of Biological Sciencesclos
?????? ??????????????? ???????????? ??? ????????? ????????? ??? ??????
Get PDFDepartment of PhysicsSophisticated longitudinal phase space (LPS) manipulations are getting attention in modern accelerator facilities. In X-ray free electron laser, temporal resolution of X-ray pulse is determined by electron beam's bunch length. To achieve temporal resolution of femtoseconds (fs) level, the beam necessitates strong bunch compression for bunch length of fs level. Also ultrafast electron diffraction requires fs level of bunch length for probe of time-resolved crystallography. In advanced accelerator concept, beam-driven wakefield acceleration uses drive and witness beams. The drive beam induces wakefield along plasma or dielectric structure and the following witness beam experiences this wakefield. Since the drive beam generates the wakefield, the gradient and transformer ratio of the acceleration are determined by its peak current and current distribution. Bunch compression and current profile shaping could provide high gradient and transformer ratio.
Conventional LPS manipulation can be divided into three partsenergy control, bunch compression, and current profile shaping. The energy control is normally performed with radio-frequency (rf) cavity or wakefield structure. Both have longitudinal electric field so that the beam can change its energy. The bunch compression has two typerf bunch compression and magnetic bunch compression. The rf bunch compression is effective at low energy beam because high energy particles catch up low energy particles during drift. The magnetic bunch compression imparts different path length depending on energy. This is effective at high energy beam because rf bunching regime is not working. The current profile shaping has two ways. First one is ultraviolet (UV) laser pulse shaping using optical devices because electron beam from a photocathode follows UV temporal distribution. Second method is direct beam control using dispersion and transverse manipulation. Although the LPS manipulation techniques are continuously developed, it presents several challenges.
Emittance exchange (EEX) is a technique that exchanges transverse and longitudinal phase spaces. This means EEX enables to convert transversely manipulated phase space to LPS. Comparing to directly longitudinal manipulation, transverse manipulation has diverse degree of freedom. However, EEX has two issuesemittance and jitters. Electron beam from a photocathode gun has large longitudinal emittance comparing to horizontal one. EEX makes large horizontal emittance which is fatal to accelerators. Also timing and energy jitters are converted into horizontal jitters that make difficulties on transverse beam control.
Double emittance exchanged (DEEX) returns these things to original direction and also performs LPS manipulation. EEX provides transverse to longitudinal phase space exchange, hence transverse manipulation in between two EEXs enable to control longitudinal phase space. The middle section in between two EEXs is composed of magnets, and these allow the precise and simple transverse manipulation. In this thesis, we present demonstration results of DEEX beamline with three functions (tunable bunch compression, longitudinal chirp control, and nonlinear phase space manipulation), and its feasible applications. These functions work in a flexible manner and contain diverse opportunities on accelerator applications.ope
Nanoscale analysis of catalytic reaction on the surface of metal and semiconductor
No full textDepartment of Materials Science and Engineeringclos
????????? ????????? pH ????????? ???????????? ???????????? ??????
No full textSchool of Energy and Chemical Engineering (Chemical Engineering)clos
Prototyping as a Qualitative Inquiry in Fuzzy Front-End of New Service Development
Get PDFGraduate School of Creative Design Engineering DepaPrototyping has been an ever-evolving activity in design with the expansion of the design field ??? from physical objects to software and then towards service and experience. In service design, prototyping represents the amalgamation of different aspects ??? i.e., forms, purposes, where it is situated in the design process, and the context of use. Despite the abundance of literature on prototyping in design, how prototyping can help novice designers validate service concepts for the real-life context in the front-end needs more attention.
This thesis aims to shed light on how prototyping helps novice designers and clients mitigate the vagueness in front-end service concept design, specifically focusing on the explicit role of engaging (potential) users.
By revisiting three service concept design projects that included user workshops utilizing prototypes, semi-structured interviews were undertaken with six student designers and three client personnel with visual stimuli as a reminder of the workshops. Inductive thematic analysis unveiled how the quality of conversation between designers and (potential) users varied by the intentionally adjusted attributes in prototyping. Additionally, the result identified the situated challenges and implications of prototyping in different organizational contexts. The findings show that the ambiguity intrinsic to the front-end design is challenging for novice designers, which later gets resolved by engaging with (potential) users, refining service concepts, and having a consensus with clients.
I hope this thesis suggests a practical guide for novice designers in realizing their design intention into prototyping activity to conjugate the fuzzy front-end as a space for exploring design directions.ope
Sequential Process-Based Perovskite Solar Cells toward Large-Area Application
No full textDepartment of Materials Science and Engineeringclos
Spatiotemporal Chaoticity Control for Solving Combinatorial Optimization in Memristor Neural Networks
No full textGraduate School of Artificial IntelligenceRecently, artificial neural networks are re-visited as an efficient solution to the combinatorial optimization problem, which is computationally intractable, thanks to the parallel operation based on the memristor crossbar array. However, the high convergence of the artificial neural network based on gradient descent also has limitations in solving large-scale combinatorial optimization problems because it is not possible to confirm whether it is an actual optimal solution without exploring all possible cases. Therefore, the conventional approaches to solving the problems so far are largely divided into two methodsa deterministic method exploring all the possible cases but bound to reduce the invalid cases and a stochastic method converging to a valid solution efficiently but bound to search for different solutions as many as possible. Since the stochasticity has been temporal controlled ignoring spatial anisotropy in solution space, the spatiotemporal stochasticity control can be effective especially when the global optimum is isolated from the local optimum by excluding redundant exploration around the local optimum. It was verified in a stochastic Hopfield neural network model based on the spontaneous conductance decaying characteristics of memristors. This study envisages the high potential of semantic elaboration to exploit in-memory processing in a memristor crossbar array.ope