Does construct overload truly overload the performance? : an experimental study of experienced data modeler

서지혜 Seoul National University 2017 국내박사

A principal activity in information systems development involves building a conceptual model of domain that an information system is intended to support. Such models are created using a conceptual-modeling grammar fundamental means to specifying information systems requirement. However, the actual usage of grammar is poorly understood and some issues regarding conceptual grammar such as construct overload still remain unsolved. With regard to construct overload in conceptual modeling, past studies have had some deficiencies in research methods and even have presented contradicting results. In this paper, we experimented to test whether construct overload enables conceptual models users to understand a domain more efficiently. To acquire a more complete and accurate understanding of construct overload, our study focused on three major points; the evaluation of conceptual modeling grammar semantics, research participants and domain familiarity. This paper’s key contribution is that it is one of the first studies to investigate practitioner’s aspects of construct overload employing different degrees of domain familiarity by investigating the cognitive processes of practitioner. In addition, this research reconciles conflicting outcomes by examining practical directions for model variation. The result of study will broaden the perspective on usability in the context of the conceptual model and may serve as an ontological guidance to construct overload when modelers create a conceptual model.

Construction of hydrogeological conceptual model for analysis of groundwater seepage from the coastal landfill

Juhee Kim 고려대학교 대학원 2023 국내석사

내륙의 제한된 토지 이용 가능성으로 인해 폐기물을 매립하는 곳으로 해안 지역 및 해안 간척지가 고려되고 있다. 해안 매립장을 비롯하여 다양한 지하개발을 위해 대심도 굴착이 수행될 경우, 일반적으로 지하수의 유출이 발생한다. 본 연구는 해안 매립지 공사 중 누수 발생에 대한 사례가 보고된 석문 간척지(충청북도, 대한민국)를 대상으로 수치적모델링을 통해 공사부지내 지하수 유출의 원인을 규명하고 배수전략을 수립하였다. 연구지역에서는 20m 까지 대심도 굴착이 수행된 공사부지에서 3 개 지점의 보일링과 하루 900 톤의 누수가 관측되었다. 50 여개의 시추공을 해석하여 퇴적층, 풍화토층, 기반암으로 구성된 지질모델을 구축하였다. 수리지질학적 개념모델을 석문매립장에서 발생된 고유량의 유출 및 높은 수위를 유지함에 따라 굴착면을 통해 스스로 물이 흘러나오는 현상의 설명 가능성을 고려하여 결정하였다. 수리지질학적 개념모델을 기반으로 부정류 지하수 모델을 구축하여 유출량에 대한 정량적 해석을 수행하였다. 부지 내 누수는 -22 El.m. 하부에 위치하는 고투수성 풍화토층과 파쇄기반암을 통한 지하수의 유출로 설명할 수 있으며, 고유량 유출은 보일링 현상에 의한 지하의 고투수성 매체와 연결성으로 인해 발생되는 것으로 해석할 수 있다. 단계적으로 추가 굴착 시, 전반적으로 유출량이 증가하는 것으로 보이나 기반암이 노출되는 시점에서는 수리지질학적 개념모델에서 작용하는 주 대수층에 따라 유출량 변화 양상의 차이가 예상되었다. 수리지질학적 개념모델을 토대로, 추가 굴착 공사를 진행하기 위한 배수 계획 및 추가 굴착 후 감압정을 이용한 매립장 관리에 대한 모델링을 수행하였다. 1000 톤/일 감압정 운영을 통해서는 기존 유출량의 30~40%를 줄일 수 있는 것으로 예측되었다. Coastal areas and coastal reclaimed land are being considered as landfills for waste due to limited land availability of inland. When deep excavation is performed for various underground developments, including coastal landfills, groundwater seepage generally occurs. In this study, the cause of groundwater seepage within the construction site was identified and drainage strategy was established through numerical modeling on the Seongmun reclaimed land (Chungcheongbuk-do, Republic of Korea) where leaks were reported during coastal landfill construction. In the study area, boiling at three points and leakage of 900 tons per day were observed at the construction site where deep excavation was performed up to 20 m. By analyzing about 50 boreholes, a site geological model composed of sedimentary layers, weathered soil layer, and weathered rock layer was constructed. The hydrogeological conceptual model was determined by considering the possibility of explaining the high flux discharge and the phenomenon of water flowing out through the excavation surface by itself as the water level was maintained high. Quantitative analysis of groundwater seepage was performed by establishing a transient groundwater model based on the hydrogeological conceptual model. Leakage within the site can be explained by the groundwater seepage through the crushed bedrock and the highly permeable weathered soil layer located below -22 El m. The high flux discharge can be interpreted as occurring due to the connection with the underground high-permeability medium caused by the boiling phenomenon. When additional excavation is performed step by step, seepage seems to increase overall, but at the time bedrock is exposed, differences in seepage change patterns are expected depending on the main aquifer acting in each hydrogeological conceptual model. Based on the hydrogeological conceptual model, the drainage plan for further excavation and the management of the landfill using the relief well after additional excavation were modeled. It was predicted that 30-40% of the existing seepage could be reduced through the relief wells operation of 1000 tons/day.

Generic System Design Methodology for eVTOL Aircraft using Enhanced Electric Propulsion Modeling

이동욱 서울대학교 대학원 2023 국내박사

In recent years, various technological advances in electrification, automation, and vertical takeoff and landing (VTOL) have matured enough to enable innovation in urban aviation, resulting in the emergence of a new air transportation system known as advanced air mobility (AAM). VTOL aircraft for AAM is powered by an electric propulsion system, which provides new design freedom for the configuration and flight mechanism. To design the eVTOL aircraft for AAM, therefore, it is necessary to have generality rather than being limited to a specific concept. In addition, since eVTOL aircraft for AAM services between the intracity and intercity, the noise generated by the rotary-wing system should be considered a performance indicator for the eVTOL aircraft design. Meanwhile, due to the low specific energy of the current battery technology, the range of most full-EPS (FEPS) powered VTOL aircraft has been so far restricted to intracity operation. However, as the battery technology matures, VTOL aircraft for AAM will likely use a FEPS powered only by batteries. This requires sophisticated EPS modeling techniques to consider the electrical characteristics of each electrical device in a more accurate manner. To this end, this study proposes a generic design methodology that considers eVTOL AAM vehicles' characteristics. This design methodology comprises five modules (flight analysis, propulsion system sizing, mission analysis, weight estimation, and noise prediction) that can consider the diversity of configurations, flight mechanisms, EPS architectures, and performance assessment, including noise prediction. First, the comprehensive flight-analysis module is created by assembling component-analysis methods, including the shrouded rotor and distributed propulsor (DP). The proposed technique allows for analysis of the configurations and flight mechanisms of various types of VTOL aircraft—wingless, vectored-thrust, and lift+cruise. In addition, the scope of propulsion system sizing, mission analysis, and weight estimation has been expanded to include not only FEPS but also various Hybrid-EPSs (series, parallel, and series-parallel). Using the Farassat 1A formulation with compact loading models, it is possible to predict the thickness and load noise of eVTOL aircraft at the conceptual design stage. Also, this study proposes novel enhancements to the EPS modeling approach for FEPS-powered VTOL aircraft conceptual design by considering the electrical characteristics of each electrical device in a more accurate manner. To this end, three modules for motor, inverter, and battery analysis are constructed using equivalent circuits and semi-empirical models. First, the motor analysis module is developed using equivalent circuit analysis with the operation control strategy for a permanent-magnet synchronous motor. Second, the inverter analysis module is built using average loss models for the switching and conduction losses. Third, the battery analysis module is improved using the near-linear discharge model to consider the voltage drop during operation. Moreover, additional modules, such as those for calculating the battery stack in series and parallel, as well as regression models for the motor and inverter parameters and consideration of the drive system type (direct or indirect, including a reduction gear), are implemented. In addition, three types of applications are performed to demonstrate the necessity and capability of the proposed numerical methods for eVTOL conceptual design. In the first application, a comparative study is performed to demonstrate performance variations after replacing a FEPS with a HEPS; it is shown that HEPS with an optimal hybridization ratio has overwhelming superiority regarding payload capacity and mission range over FEPS based on the current battery technology level. In the second application, it is confirmed that changes in the type of drive system (direct or indirect) and gear ratio significantly impacted EPSs' efficiency and size, which can only be considered in the enhanced EPS approach. Lastly, the final application is to investigate the influence of noise prediction on the design optimization of an eVTOL aircraft. It is identified that there is a tradeoff relationship between noise mitigation and gross weight minimization depending on the rotor's torque and rotational speed. 항공기 전동화와 더불어, 항법 시스템 및 수직이착륙 기술의 발전은 새로운 항공 운송시스템인 차세대 항공교통(Advanced Air Mobility) 이라고 하는 항공 운송 시스템의 혁신을 가능케 하였다. 현재 700여종 이상의 차세대 항공교통용 전기동력 수직이착륙 항공기가 개발 중인 것으로 알려지고 있으며, 이들은 전기동력 기반 추진시스템을 이용함에 따라 얻어지는 높은 설계 자유도에 의해 다양한 수직이착륙 항공기 형상, 비행 메커니즘을 가진다. 즉, 차세대 항공용 수직이착륙 항공기를 설계하기 위해선 특정 컨셉에 국한되지 않는 범용성을 갖춰야 한다. 또한 도심을 비롯해 지역 거점 간 비행하기 때문에 회전익 시스템으로부터 발생하는 소음 또한 설계의 성능 지표로 활용되어야 한다. 한편, 현재 배터리 기술이 갖는 한계에 의해, 차세대 항공교통용 전기동력 수직이착륙 항공기는 도심 주변 비행 용도로 완전 전기추진방식을, 지역 거점 간 비행 용도로 하이브리드 전기추진방식을 채택하고 있다. 향후 배터리 기술이 충분히 발전한다면, 완전 전기추진방식은 탄소중립 실현을 위해 모든 차세대 항공교통 비행 임무용도로 채택될 것으로 예상되며, 이에 따라 개념설계 단계에서부터 모터의 위상각 제어, 배터리 전압 강하와 같은 전기적 특성을 고려할 수 있는 모델링에 대한 필요성이 대두되고 있다. 본 연구에서는 차세대 항공교통용 전기동력 수직이착륙 항공기만의 특징들을 고려할 수 있는 범용적인 설계 기법을 제안하였다. 제시된 설계 기법은 항공기 형상, 비행 메커니즘 및 전기추진시스템의 다양성과 소음 예측을 포함한 성능 평가를 고려할 수 있는 5개의 새로운 모듈(비행 분석, 추진 시스템 사이징, 임무 해석, 중량 추정, 소음 예측)로 구성된다. 먼저, 기존의 고정익 및 회전익 연구들에서 제시한 날개, 틸팅시스템, 덕티드 로터 및 분산추진장치등에 대한 해석 기법들을 하나의 비행 해석 모듈로 결합하여, 해석 대상의 범용성을 높였다. 또한, 추진 시스템 사이징, 임무 해석 및 중량 추정 대상의 범위를 완전 전기추진방식뿐만 아니라 하이브리드 전기추진방식(직렬, 병렬, 그리고 직병렬)까지 포함할 수 있도록 확장하였다. 그리고 음향상사식 Farassat 1A formulation을 이용하여 특정 고도에서의 두께 및 하중 소음을 예측할 수 있도록 하였다. 또한 본 연구에서는 완전 전기추진방식을 이루는 각 전기장치 (전기모터, 인버터, 배터리)의 전기적 특성을 보다 정확하게 고려할 수 있는 새로운 전기추진시스템 모델링 방법을 제안한다. 제시된 전기추진시스템 모델링 방법은 영구자석 전기모터의 MTPA (Maximum Torque Per Ampere) 제어 전략이 접목된 등가회로 모델, 인버터의 스위칭 및 전도 손실 모델, 상용 배터리 방전 그래프의 선형화 모델 등으로 구성된다. 그리고 모터 및 인버터 매개변수에 대한 회귀 모델을 추가적으로 구축하여, 새로운 전기추진시스템 모델링 방법을 개념 설계 기법에 적용 가능하게 하였다. 차세대 항공교통용 전기동력 수직이착륙 항공기 개념 설계를 위해, 본 연구에서 제안된 수치 기법들의 필요성과 기능을 입증하기 위해 세 가지 유형의 응용 연구를 수행하였다. 첫 번째 응용 연구는 완전 전기추진방식과 하이브리드 전기추진방식과의 비교 연구로, 최적의 하이브리드화 비율을 구현한다면 비행 거리를 완전 전기추진방식 대비 3배 가량 확장시킬 수 있음을 보였다. 두 번째 응용 연구는 전기추진시스템 모델링 방법 간의 차이점과 완전 전기추진방식의 모터 구동 시스템(직접 혹은 간접) 및 기어비의 변화가 전기추진시스템의 효율 및 사이즈에 미치는 영향을 확인하였다. 마지막 응용 연구로, 소음 저감이라는 설계 요구조건이 전기동력 수직이착륙 항공기 설계에 미치는 영향에 대해 알아보았으며, 소음과 항공기 총 중량은 서로 명확한 반비례 관계를 가지는 것을 확인하였다.

Construction of hydrogeological conceptual model and groundwater model in standard watershed-scale

김민경 Graduate School, Korea University 2022 국내석사

To establish a water resource management plan considering groundwater and surface water, analysis of the watershed-scale groundwater reserve and budget should be preceded. In this research, the watershed-scale hydrogeological conceptual models (HCMs) and numerical groundwater model were constructed using public data on subsurface environment. The study was conducted in the lower watershed of Paldang Dam located (Gyeonggi-do, South Korea). The public data on the surface, topography, hydrology, and geology of the research area were obtained from public database in South Korea. From the borehole log in the watershed, ten types of geological media were observed, and five stratigraphic units were defined based on the lithologic characteristics. Stratigraphy modeling was performed using the strong relationship between surface elevation and the appearance depths of stratigraphic units. The characteristics of the depth of appearance in the geological media were different in the area around the Han River, in the mountainous area, in the Wolmuncheon area. HCMs were constructed by defining hydrogeological units through subgrouping of lithologic units to present the hydraulic connectivity between the stratigraphic units. The steady-state groundwater flow was numerically simulated using MODFLOW and conceptual model approach by reflecting the hydrogeological conceptual models. The aquifer parameters of the hydrogeological units were determined using PEST so that the groundwater model could simulate the groundwater levels at 52 observation points. Validity of the HCM and the groundwater model of the watershed were compared using R2 and sum of squared error of the optimized groundwater models for the different HCMs. Results have shown that the HCMs of multi-layer models were better than the single layer model. The soil group and fractured rock group should be separated in HCM due to the huge difference of hydraulic conductivity. The confining layer between the soil group and fractured rock group cannot be used to improve the model fit, which means the hydrogeological connection of water bodies in soil and fractured rock. Separating the weathered soil layer from the soil group did not improve the fit of the model either. Separation of intact bedrock (Hard rock) layer from fractured rock group did not improve the fit of the model, either, but this was just due to the lack of observations in the deep subsurface environment. For a reliable evaluation of groundwater resources, it is necessary to accurately reflect the storativity of the intact bedrock layer. Based on the volume and porosity (or storativity) of the hydrogeological unit of the constructed model, the groundwater reserves were estimated to be 2.163E+12 m3. For the reliable evaluation of groundwater resources, it is necessary to improve the characterization technique for the volume and storativity of the bedrock.

Developing an conceptual integrated model and measurement for the employee’s learning agility

이진주 한양대학교 대학원 2018 국내석사

(A) study on the effects of weather forecasters' cognitive biases and utilization of conceptual models on decision making

연상훈 Graduate School, Yonsei University 2023 국내박사

본 연구는 기상예보관의 의사결정에 영향을 미치는 직관적인 요인에 대하여 인지편향의 발생과 정신모형 즉, 기상현상에 대한 입체적인 개념모형의 활용을 중심으로 수행되었다. 단기경험(예보빗나감 또는 예보놓침)을 겪은 이후에도 외부 비판 압박감으로 인해 악기상 경보 발령은 후회를 최소화하는 방향으로 편향될 것이라는 가설(1)에 대한 조사에서는 예보놓침으로 인한 평판 또는 신뢰저하 위험에 대한 우려가 가장 큰 것으로 나타났으며, 따라서 예보놓침의 후회를 최소화하는 방향으로 결정이 편향되는 결과를 보였다. 모호한 기상상황에서의 잘못된 판단에 대한 두려움에 대처하기 위해 인근 기상대가 악기상 경보를 발령하면 그 결정을 따르는 편향이 나타날 것이라는 가설(2)에 대한 조사에서도 인근 기상대가 악기상 경보를 발령하는 상황에서 동일한 결정을 하는 편향을 보여주었다. 다만, 인근 기상대가 악기상 경보를 발령하지 않는 상황에서는 가설(1)에 의해 예보놓침에 대한 후회 최소화 방향으로의 편향이 우세할 것으로 추정되었으나, 발령과 미발령 결정이 거의 절반씩을 차지하는 것으로 보아 인근 기상대의 미발령 상황도 역시 판단에 영향을 주는 것으로 보아야 할 것이다. 악기상 경보가 적중하였다 하더라도 악기상에 의한 피해의 정도에 따라 자기평판에 대한 평가는 편향될 것이라는 가설(3)에 대한 조사에서는 피해가 극심한 경우 자신 또는 팀에 대한 평판이 나빠졌을 것이라고 평가하는 경향이 우세하였다. 악기상 경보가 적중하였는데도, 자신 또는 팀에 대한 평판 평가는 나빠졌을 것이라고 보지만 의사결정 과정에 대해서는 피해가 극심한 경우나 미미한 경우 모두에서 좋은 평가를 내렸다. 3가지의 가설에 대한 분석 결과, 기상예보관은 악기상 경보를 발표하는 상황에서 자신 또는 팀에 대한 평판 저하의 위험성을 인식하고 있음을 실증할 수 있었다. 기상예보 과정에서 기상현상을 설명하는 3차원 개념모형의 활용 빈도가 높을수록 예보적중률에 긍정적인 영향을 줄 것이라는 가설(4)에 대한 조사에서는 정신모형의 활용(mental representation)은 Supercell storm의 예측 적중과 가장 높은 양의 상관관계를 보였다. 이는 선행연구의 타당성을 확인하는 차원의 결과이기도 하지만, 선행연구에서 다루지 않았던 예측적중과의 직접적인 상관관계를 알아보는 것은 첫 시도였다는 점에서 강조되어야 한다. 수치예보 강수예측성능 해석시 휴리스틱을 유발할 수 있는 오차는 표본크기가 작은 강수유형에서 발생할 것이라는 가설(5)에 대한 조사에서는 표본크기가 가장 작았던 대류성 강수가 검증지수 방법에 따라 검증값의 편차가 크게 나타나고 기준점과 조정 휴리스틱을 유발할 수 있는 상황에 가장 가까운 것으로 추정되었다. 따라서, 특정 기간에 대하여 강수라는 기상요소 하나에 대한 일반적인 검증보다는 강수 발생 메커니즘에 따라 분류한 강수유형별로 검증을 수행하여 수치예보모델 검증의 표본크기 민감도를 고려해야 할 것이다. 본 연구를 통해서 국가와 사회 방재시스템의 일원으로서 중요한 의사결정을 해야하는 기상예보관에게 발생할 수 있는 인지편향을 알아보았고, 지식과 경험의 누적으로 얻어지는 직관적인 개념모형 활용의 긍정적인 요인에 대해서도 분석할 수 있었다. 이 연구의 결과를 바탕으로 다음과 같은 방안들을 제안하였다. 첫째, 탈인지편향 방안이다. 우선 현업기관별 정규 기상예보관 교육과정에 탈인지편향 커리귤럼을 반영하여 직접적인 처방을 하는 것이다. 또한 현업 교대근무 사후분석서(after action review) 작성을 절차화 하는 것이다. 기상예보 현업과정의 대부분은 암묵지(tacit knowledge) 형태로 전이되고 있는데 사후분석서 작성을 통해 암묵지의 명시지화를 추구하는 방안을 제시하였다. 또한, 기상예보시 노이즈(noise)를 제거하고 시그널(signal)을 강조하는 앙상블 확률 예측을 바탕으로 기상-방재 공동의사결정시스템을 구축하는 것이다. 기상학적 기준으로만 악기상 경보를 발령하는 것이 아니라 방재분야에서도 악기상에 의한 피해액, 예방조치에 의한 비용 추정치와 기상예보시스템의 예측성능 관련 수치를 대입하여 경제적 가치를 산출하는 모델을 활용하는 방안이 될 것이다. 둘째, 개념모형 활용 능력 강화 방안이다. 기상학 교육을 개념모형 구축과 실제 사례연구에 중점을 두고 커리귤럼을 구성하여 운영 중인 미국 UCAR의 COMET 프로그램을 벤치마킹하여 능동적인 제작으로의 교육정책 전환이 필요하다. 또한 방대한 기상자료를 사례별 저장소로 분류하고 기상시뮬레이터(weather simulator)를 개발하여 무제한 시연이 가능하도록 하는 것이다. 이것은 기상예보관이 경험을 누적하는데 극복해야 하는 시공간 제약을 해소하는 효과를 거둘 수 있다. 이와 더불어 기상예보관의 눈 움직임 추적(eye-tracking)과 상황인식 정도를 평가한 연구를 바탕으로 본 연구의 한계를 넘을 수 있을 것으로 보이며, 탈인지편향의 근원적 방안으로서 인공지능에 의한 악기상 예보 추론과정 개발을 추진하는 것이 본 연구에서 드러난 한계를 극복하고 미래지향적인 연구의 방향으로 삼을 수 있을 것이다. This study focused on cognitive bias and using a mental model, a three-dimensional conceptual model for meteorological phenomena, for intuitive factors that affect weather forecasters' decision-making. In an investigation of hypothesis (1) that the issuance of a severe weather warning would be biased in the direction of minimizing regret due to the pressure of external criticism even after experiencing a short-term experience (a false positive or a false negative), concern about the risk of reputation or trust loss due to false negatives was found to be the greatest. Thus the decision was biased toward minimizing the regret of a false negative situation. In an investigation of hypothesis (2) that if a nearby weather station issues a severe weather warning to cope with the fear of wrong judgment in ambiguous weather conditions, a bias will follow that decision, and the same decision will be made in a situation where a nearby weather station issues a severe weather warning. However, in a situation where a nearby weather station does not issue a severe weather warning, it was estimated that the bias in minimizing regret for false negatives would prevail according to hypothesis (1). Therefore, the situation of non-issuance should also be seen as affecting the judgment. Finally, in the investigation of hypothesis (3) that the evaluation of self-reputation would be biased according to the degree of damage caused by the severe weather even if the severe weather warning was hit, the tendency to evaluate that the reputation of oneself or the team would have deteriorated if the damage was extreme. However, the decision-making process was evaluated as good when the damage was extreme or insignificant. As a result of the analysis of the three hypotheses demonstrated that weather forecasters are aware of the risk of reputational damage to themselves or their teams in the situation of issuing severe weather warnings. In the investigation of hypothesis (4), the higher the frequency of use of the three-dimensional conceptual model that explains meteorological phenomena in the weather forecasting process, the more positive the impact on the forecast accuracy rate showed the highest positive correlation. Although this confirms the validity of the previous study, it should be emphasized that it was the first attempt to determine the direct correlation with the predicted hit, which should have been covered in the previous study. In the investigation of hypothesis (5) that the error that can cause heuristics in the interpretation of NWP(Numerical Weather Prediction) performance will occur in precipitation types with a small sample size, convective precipitation with the smallest sample size is the deviation of the verification value according to the verification index method. Therefore, rather than a general verification of one meteorological factor, precipitation, for a specific period, verification should be performed for each precipitation type classified according to the precipitation mechanism to consider the sample size sensitivity of the NWP model verification. Through this study, the cognitive bias that can occur in weather forecasters who have to make crucial decisions as a member of the national and social disaster prevention system was investigated, and the positive factors of using an intuitive conceptual model obtained through the accumulation of knowledge and experience could also be analyzed. Finally, based on the results of this study, the following measures were proposed: (1) It is a de-biasing method. The direct prescription is made by reflecting the post-cognitive bias curriculum in each regular weather forecaster training course. (2) In addition, it is procedural to prepare after-action reviews for work shifts. Most weather forecasting work-related processes are being transferred to the form of tacit knowledge, and a plan to pursue the explicit knowledge of tacit knowledge through post-analysis was proposed. (3) In addition, it is to build a weather-disaster prevention joint decision-making system based on ensemble probability prediction that removes noise and emphasizes signals in weather forecasting. Not only issuing severe weather warnings based on meteorological criteria but also in the field of disaster prevention, it will be a plan to utilize a model that calculates economic value by substituting the estimated cost of damage caused by severe weather, preventive measures, and the predictive performance of the weather forecasting system. Next is a plan to strengthen the ability to use the conceptual model. It is necessary to convert education policy to active production by benchmarking the COMET program of UCAR in the United States. It is being operated by constructing a curriculum that establishes a conceptual model and actual case studies for meteorology education. In addition, it is to classify vast meteorological data into case-by-case storage and to develop a weather simulator to allow unlimited demonstrations. Finally, it can resolve the spatial and temporal constraints that weather forecasters must overcome in accumulating experience. In addition, based on the study that evaluated the degree of eye-tracking and situational awareness of weather forecasters, the limitations of this study can be overcome, and the reasoning process of severe weather forecast by artificial intelligence is a fundamental method of de-biasing. Promoting development can overcome the limitations revealed in this study and set a direction for future-oriented research.

온톨로지를 이용한 이산사건 시뮬레이션의 개념적 모델 구축 지원에 관한 연구

박지성 성균관대학교 일반대학원 2013 국내석사

Conceptual Modeling is the process of abstracting a model from a real or proposed system. It is probably the most important aspect of a simulation study. Relate works show that the elementary developers devoted little time to understanding how the systems actually worked, namely they didn't build appropriate conceptual model. Thus, the result of simulation is inconsistent because it depends on developer's competence. Although many researches suggested various techniques enabling developer to build conceptual model, there were several limitations. In this study, to overcome the limitations of existing techniques, we proposed COMBINE-DES(COnceptual Model BuildINg framEwork using ontology for Discrete Event Simulation). The COMBINE-DES supports expediting the conceptual modeling with Solution ontology generated by Domain ontology and Simulation ontology. Moreover, it provides consistent simulation result regardless of repeated modeling.

Impact of interpretability regularization on the transparency and inference of deep neural networks

Joo, Sunghwan Sungkyunkwan University 2025 국내박사

딥러닝 모델은 다양한 작업에서 매우 효과적이지만, 본질적인 복잡성으로 인해 해석 가능성에 큰 어려움을 겪고 있습니다. 이러한 “블랙박스” 특성은 모델이 잘못된 상관관계에 의존하여 추론할 경우 의료 진단, 대출 심사, 자율 주행과 같은 중요한 응용 분야에서 실패할 수 있는 위험을 초래합니다. 설명 가능한 인공지능은 모델을 더 투명하게 만들기 위한 방법을 개발함으로써 이러한 문제를 해결하고자 합니다. 예를 들어, 중요 영역 맵핑(saliency mapping)은 모델 예측에 중요한 입력 영역을 강조하고, 개념 기반 설명 기법은 모델의 행동을 인간이 이해할 수 있는 개념에 대한 수치로 나타냅니다. 그러나 기존의 설명가능 인공지능 방법들은 모델의 행동을 모니터링하고 수정하는 데 있어 상당한 한계를 가지고 있습니다. 일부 중요 영역 맵핑 기법은 단순한 윤곽선 감지기처럼 작동하여 모델 파라미터와 관계없이 일관된 결과를 생성하고, 일부 다른 기법은 적대적 속성 조작(AAM)에 취약한 경우가 있습니다. 특히 모델이 잘못된 또는 사회적으로 민감한 특성에 의존하는지 여부를 판별하는 데에 설명가능 인공지능이 사용된다면 이러한 취약성은 극복해야 할 큰 과제입니다. 이 논문은 해석 가능성 정규화의 영향을 조사함으로써 설명 가능 인공지능의 강건성과 실용성을 향상시키기 위한 세 가지 주요 연구 질문을 탐구합니다. 첫째, 중요 영역 기법을 속이고 실제 모델 동작을 숨기는 적대적 모델 조작(AMM) 가능성을 발견하였습니다. AMM은 모델의 악의적인 모델 개발자가 모델의 잘못된 행동을 바로잡기 보다는 우회 방법으로 테스트를 통과할 수 있는 가능성을 고려할 때 더 치명적입니다. 실험적으로, 이러한 조작은 모델 정확도에 손상을 주지 않고 최소한의 매개변수 조정만으로 달성될 수 있음을 보였습니다. 둘째, AAM에 대한 중요 영역 맵의 강건성을 높이는 연구를 수행하였습니다. 기존의 지형 평탄화 기법은 기울기 크기 정규화를 근본적으로 무시하는 한계가 존재합니다. 본 논문은 지역 기울기를 더 잘 정렬하는 새로운 정규화 접근법을 제안하였고 실험 결과 우리의 접근법이 AAM에 대항하여 강건한 중요 영역 맵을 생성하는 데 있어 기존 기술보다 우수한 성능을 보였습니다. 마지막으로, 모델 훈련에 설명 가능 인공지능을 활용하여 모델 행동을 규제하고 수정하는 새로운 방법을 소개합니다. 제안하는 방법은 상기한 바와 같이 신뢰성에 문제가 있는 중요 영역 기법 대신, 개념 기반 설명 기법을 모델 규제에 활용합니다. 비선형 개념 모델을 사용해 개별화된 개념 벡터를 얻음으로써, 단순 선형 개념 활성화 벡터(CAV) 방법보다 모델의 가짜 특징 의존도를 더 잘 낮출 수 있습니다. 실험 결과, 우리의 방법은 잘못된 상관관계 의존성을 줄이는 데 효과적이며 모델 추론의 신뢰성을 높일 수 있음을 보였습니다. 이 논문은 이론적 이해와 실질적 응용 모두에서 설명 가능 인공지능 분야에 중요한 기여를 하며, 궁극적으로 더 투명하고 신뢰성 있는 딥러닝 모델의 발전을 촉진하기 위한 방향을 제시합니다. Deep neural networks, while highly effective across numerous tasks, pose significant challenges in interpretability due to their intrinsic complexity. This “black-box” nature presents risks in critical applications such as medical diagnostics, loan processing, and autonomous driving, where reliance on spurious correlations can lead to failures when encountering distribution shift. Interpretable AI aims to addresses these challenges by developing methods that make model more transparent. For example, techniques such as saliency mapping highlights important input regions for model prediction, while concept-based approaches present model behavior to human-friendly concepts. However, existing interpretable AI face notable limitations that disturb their use for monitoring and correcting model behavior. Some saliency techniques act as simple edge detectors, producing consistent outputs regardless of model, while others are vulnerable to adversarial attribution manipulations (AAM). These susceptibilities are particularly concerning when they are used to ensure models do not rely on spurious or socially sensitive features. This thesis explores three key research questions to improve the robustness and practical use of interpretable AI, by investigating the impact of interpretability regularization. First, we focus on enhancing the robustness of saliency maps against AAM. By identifying the shortcomings of existing landscape smoothing techniques, particularly the neglect of gradient scale normalization, we propose a novel regularization approach that better aligns local gradients. Empirical results demonstrate that our approach outperforms existing techniques in generating robust saliency maps against AAM. Second, we investigate a new vulnerability known as adversarial model manipulation (AMM), where internal model parameters are modified to deceive saliency methods and conceal the true model behavior. Unlike input perturbations, AMM is more critical by considering the scenario where mean developers potentially decide to detour the model to pass the validation test using interpretable AI, rather than deliberately correct the model behavior. Empirically, we show that such manipulations can be achieved with minimal adjustments of model parameters, without hurting the accuracies. Lastly, we introduce a novel method for incorporating interpretability into model training to guide and correct model behavior. Instead of leveraging unreliable saliency maps into model regularization, we leverage concept-based methods. By employing non-linear concept head to obtain individualized concept vectors, our approach enhances conceptual sensitivity regularization, outperforming linear CAV methods in debiased learning. Empirical results demonstrate that our method better aligns model decisions with human-understandable concepts and reduces reliance on spurious correlations. This thesis makes significant contributions to the field of interpretable AI by advancing both theoretical understanding and practical applications, ultimately promoting more transparent deep learning models.

(A) study on the 3D reality model platform construction and effectiveness based on UAV aerial photos in urban development projects

이재영 Graduate School, Yonsei University 2022 국내박사

This study aims to verify the effectiveness of planning a city by establishing a 3D Reality Model platform using 3D spatial information technology to solve the problems of the use of flat technology of images based on two dimensions using existing spatial information. To this end, a work platform was designed and built to produce a 3D reality model based on large-scale aerial photographs obtained from UAVs for various objects in urban areas. The quality of the results was analyzed based on related criteria. In addition, the 3D reality model produced based on the platform was measured and analyzed through comparison between the data by the existing method and the pair by urban application item. To this end, it was designed to establish a high-quality 3D reality model considering specificities such as topography and facilities in the urban area and various environmental factors in the urban area. This study is to analyzes the effectiveness of the 3D reality model. In order to use the produced 3D reality model for urban development projects, the accuracy verification and quality evaluation of the 3D reality model must be preceded. Therefore, the evaluation of the 3D reality model was performed by classifying it into a quantitative evaluation to evaluate location accuracy in the quality evaluation criteria for building 3D national spatial information and a qualitative evaluation to evaluate the consistency of detail level and visualization of the object to be studied. As a result of the evaluation, it was judged that the location accuracy, detail level, and visualization level were secured to meet the relevant regulations, thereby proving the possibility of using the 3D reality model throughout the urban development project. In this study, six evaluation factors such as presence, spatial ability, conceptual understanding, aesthetic, work efficiency, and reliability were set to analyze effectiveness after platform construction. In addition, an expert survey was conducted as an evaluation method to analyze the effectiveness, and statistical analysis was conducted on six effectiveness evaluation factors. A comparison was made between the existing 2D method before the platform was built and the 3D method after the platform was built. To this end, in this study, the differences between the elements were analyzed in which items. Furthermore, one group pre-post test design was applied to analyze this objectively. In addition, a survey was conducted on the same person before and after the platform was built, and the survey method was conducted through an electronic survey provided by Google through an online Google Form survey. As a result of the survey, 183 copies of the questionnaire were collected, and 90.6% of the respondents showed a preference for the effectiveness of the 3D methodology through platform construction. In particular, the response rate to the 3D methodology in terms of presence, spatial ability, work efficiency, and reliability is relatively high, so the realistic judgment on the space of the 3D methodology, the resulting increase in work efficiency, and the effect on data reliability are favorable. Based on the above results, it was possible to verify the effectiveness of the platform and the 3D reality model for creating a 3D reality model in terms of understanding the space for the basic investigation and business efficiency. In addition, as a result of conducting a paired T-test to verify the difference between the existing method before building the platform and the 3D application method after building the platform by applying the pre-post inspection design method for a single group, all six factors showed negative t values below the statistical significance level. All of the survey items were found to have significant differences and were higher in the 3D method than in the 2D method. As a result of questioning and analyzing other opinions on ways that UAV photos can be used in the field of work to ask for necessary technical factors in the future, many respondents said they were related to the status investigation and impact analysis. In addition, there were opinions on urban and building changes, record management, landscape analysis, and utilization of construction site construction, process, and safety management aspects in the development area. Other and minority opinions suggested using tourism, travel, and educational services and the introduction of metaverse implementing artificial intelligence and VR. In summary, many opinions on the current status investigation analysis and urban management aspects were presented, and through other opinions, future technology applicability, feasibility, and research tasks were examined, and qualitative opinions that were not covered in the quantitative evaluation of the previous effectiveness evaluation were indirectly collected. In the future, if the technology examined in this study is combined with technologies such as metaverse, artificial intelligence, BIM, and Digital Twin, it is expected that the built spatial information data will be able to analyze and utilized whenever desired. On the other hand, active measures are needed to consider the problems caused by restrictions by laws or regulations and the reality that they are not yet applicable, such as lack of awareness of 3D technologies.

Generative Neuro-Symbolic Models of Concept Learning

Feinman, Reuben New York University ProQuest Dissertations & These 2023 해외박사(DDOD)

소속기관이 구독 중이 아닌 경우 오후 4시부터 익일 오전 9시까지 원문보기가 가능합니다.

Human concepts exhibit a collection of unique qualities that together are not well-explained by current computational methods. On the one hand, human conceptual knowledge is distinguished for its productivity and generality: people learn new concepts very quickly from just one or a few examples, and the representations they acquire can be applied flexibly to a variety of tasks without retraining. In addition, people's conceptual knowledge interacts directly with raw signals: people learn new concepts directly from raw, high-dimensional data, capturing complex correlations and invariances that support the recognition and generation of new examples in equally high-dimensional media. Two modeling traditions have explained different components of these empirical phenomena, but we lack a unified computational framework to understand and account for the collective capabilities.This thesis presents a new computational framework for modeling human concepts that builds on two rich traditions in cognitive science. We hypothesize that human concept representations include a combination of structural and statistical ingredients, and that models with an appropriate synthesis of these ingredients will help account for the collective capabilities of human concept learning. Our approach---dubbed Generative Neuro-Symbolic (GNS) modeling---uses the control flow of a probabilistic program, coupled with symbolic primitives and renderers, to model the causal and compositional processes by which concepts are formed. At the same time, it integrates neural network subroutines to interface directly with raw data and capture complex correlations in observations. We demonstrate two instances of this approach developed to model human concepts of handwritten characters and synthetic "alien figures." Our experiments show that GNS provides a useful framework to understand the dual structural and statistical natures of human concepts and account for a diversity of capabilities. Additional experiments explore alternate ways to integrate structural and statistical representation and account for psychological phenomena, studying the dynamics of learning-to-learn and the acquisition of inductive biases in neural network models.