본문

서브메뉴

Designing From End-to-End and Learning Control Policies on the Edge With Data-Driven Optimization: Applications to Adaptive Plasma Medicine.
Designing From End-to-End and Learning Control Policies on the Edge With Data-Driven Optim...
Designing From End-to-End and Learning Control Policies on the Edge With Data-Driven Optimization: Applications to Adaptive Plasma Medicine.

상세정보

Material Type  
 학위논문
 
0017161775
Date and Time of Latest Transaction  
20250211151443
ISBN  
9798384449386
DDC  
660
Author  
Chan, Kimberly J.
Title/Author  
Designing From End-to-End and Learning Control Policies on the Edge With Data-Driven Optimization: Applications to Adaptive Plasma Medicine.
Publish Info  
[S.l.] : University of California, Berkeley., 2024
Publish Info  
Ann Arbor : ProQuest Dissertations & Theses, 2024
Material Info  
172 p.
General Note  
Source: Dissertations Abstracts International, Volume: 86-03, Section: B.
General Note  
Advisor: Mesbah, Ali.
학위논문주기  
Thesis (Ph.D.)--University of California, Berkeley, 2024.
Abstracts/Etc  
요약Cold atmospheric plasmas (CAPs) are becoming a breakthrough technology in a variety of materials processing and characterization applications, including in plasma medicine. CAP jets (CAPJs) are a versatile tool in plasma medicine because they can be a low-cost, portable, point-of-care solution for a variety of biomedical applications. However, selecting operational parameters of CAPJs (or CAPs in general) remains an open challenge due to a variety of factors, including variability in patients (i.e., target interface), variability in CAPJ operation, sensitivity to disturbances and environmental conditions, and difficult-to-model dynamics of CAPs resulting in uncertain predictions about CAP-interface interactions. Predictive control has become the state-of-the-art in addressing aspects of the safety, reproducibility, and efficacy of CAP treatments. This dissertation addresses two open aspects of CAP control, specifically designing feasible embedded control systems for point-of-care CAPJs and designing individualized CAP treatment regimens. Together, these two aspects represent the overarching objective of this dissertation: enabling point-of-care devices for precision plasma medicine.CAPJs for biomedical applications are often touted for their portability and point-of-care use. Additionally, medicine as a field is moving towards more targeted approaches to patient healthcare due to the influx of data from personal devices (e.g., smart wearables) that track health trends and physical activity and due to the importance of considering diverse patient profiles for equitable and efficacious medical treatments. This trend (part of a tendency towards "edge computing") combined with the nonlinear, multi-variable CAP dynamics calls for embedded control policies that are capable of implementation on resource-limited hardware. The first part of this dissertation provides a novel fusion of hardware and software design (aka "hardware-software co-design") of control policies to find optimal and feasible embedded control policies on resource-limited hardware. In particular, key elements of the end-to-end design pipeline include the digital control policy, the physical computing hardware, and the closed-loop performance measures of interest such as chemical/biological effects of CAPs on target interfaces. We demonstrate that a data-driven optimization framework based on Bayesian optimization (BO), which can simultaneously incorporate the control policy design and hardware considerations when implementing the control policy, can effectively design feasible embedded control policies that target multiple objectives. An estimation of the Pareto frontier (i.e., trade-off curve) can be generated via hardware-in-the-loop simulations and used to inform the design of real-time control policies.Several applications in plasma medicine require repeated treatments to realize therapeutically effective treatment outcomes to avoid overdosing and/or to treat long-term conditions. Prior works have illustrated predictive control strategies are capable of safely delivering CAP treatments to patients, but these strategies generally rely on underlying assumptions of individual subject characteristics (i.e., empirical models based on population data). This consideration necessitates adaptive treatments that are updated via observations of treatment outcomes, which can be addressed through data-driven optimization. In simulations and experiments, we demonstrated that deep learning-based control policies, which are amenable to resource-limited hardware, can be updated directly using multi-objective BO. We demonstrated how deep learning-based control policies can be updated to find the optimal trade-offs in treatment objectives when characteristics of individual subjects may differ from the population. In a complementary direction, we developed a novel strategy to safely explore the individualized objective space without compromising on performance improvements. We demonstrated that our safe explorative BO strategy finds a balance between overly-cautious exploration that may get stuck at local optima and overly-eager exploration that may violate safety-critical constraints.The primary focus of this dissertation was on the therapeutic benefits of CAPs. The final contribution of this dissertation investigated a novel aspect of CAPs for biomedical use: (biological) material characterization. We demonstrated that CAPs are uniquely capable of producing minimally destructive effects during interactions with biological tissues that can be used to identify and classify different tissue types. A key aspect of this finding is that real-time chemical and electrical measurements of plasma-tissue interactions can be analyzed in physics-informed ways and fed into machine learning strategies to predict the type of a biological tissue. Results from this study can have significant implications in non-invasive early skin cancer detection systems and/or in real-time surgical assistance.To conclude, this dissertation presented results that illustrate an end-to-end journey from the design of physical computing hardware to the design of digital control policies to the design and characterization of (bio)chemical outcomes of plasma treatments in medicine. This dissertation established that data-driven optimization is a versatile tool to regulate and personalize the outcomes of CAP treatments. For medicine, BO mimics the doctor-patient interaction, and thus provides a natural augmentation to the medical toolkit. Future work may involve addressing additional challenges regarding connected devices and data-driven strategies (i.e., (cyber)security, privacy, distributed deployment), fusion of physics-structured learning with data, and evaluation of such methods in preclinical and clinical studies. The findings in this dissertation were grounded in plasma medicine, but can be broadly applicable to other non-equilibrium plasma applications, e.g., semiconductor processing.
Subject Added Entry-Topical Term  
Chemical engineering.
Subject Added Entry-Topical Term  
Engineering.
Subject Added Entry-Topical Term  
Information technology.
Subject Added Entry-Topical Term  
Computational chemistry.
Index Term-Uncontrolled  
Cold atmospheric plasmas
Index Term-Uncontrolled  
Data-driven optimization
Index Term-Uncontrolled  
Machine learning
Index Term-Uncontrolled  
Plasma medicine
Index Term-Uncontrolled  
Predictive control
Added Entry-Corporate Name  
University of California, Berkeley Chemical Engineering
Host Item Entry  
Dissertations Abstracts International. 86-03B.
Electronic Location and Access  
로그인을 한후 보실 수 있는 자료입니다.
Control Number  
joongbu:657476

MARC

 008250224s2024        us  ||||||||||||||c||eng  d
■001000017161775
■00520250211151443
■006m          o    d                
■007cr#unu||||||||
■020    ▼a9798384449386
■035    ▼a(MiAaPQ)AAI31296283
■040    ▼aMiAaPQ▼cMiAaPQ
■0820  ▼a660
■1001  ▼aChan,  Kimberly  J.
■24510▼aDesigning  From  End-to-End  and  Learning  Control  Policies  on  the  Edge  With  Data-Driven  Optimization:  Applications  to  Adaptive  Plasma  Medicine.
■260    ▼a[S.l.]▼bUniversity  of  California,  Berkeley.  ▼c2024
■260  1▼aAnn  Arbor▼bProQuest  Dissertations  &  Theses▼c2024
■300    ▼a172  p.
■500    ▼aSource:  Dissertations  Abstracts  International,  Volume:  86-03,  Section:  B.
■500    ▼aAdvisor:  Mesbah,  Ali.
■5021  ▼aThesis  (Ph.D.)--University  of  California,  Berkeley,  2024.
■520    ▼aCold  atmospheric  plasmas  (CAPs)  are  becoming  a  breakthrough  technology  in  a  variety  of  materials  processing  and  characterization  applications,  including  in  plasma  medicine.  CAP  jets  (CAPJs)  are  a  versatile  tool  in  plasma  medicine  because  they  can  be  a  low-cost,  portable,  point-of-care  solution  for  a  variety  of  biomedical  applications.  However,  selecting  operational  parameters  of  CAPJs  (or  CAPs  in  general)  remains  an  open  challenge  due  to  a  variety  of  factors,  including  variability  in  patients  (i.e.,  target  interface),  variability  in  CAPJ  operation,  sensitivity  to  disturbances  and  environmental  conditions,  and  difficult-to-model  dynamics  of  CAPs  resulting  in  uncertain  predictions  about  CAP-interface  interactions.  Predictive  control  has  become  the  state-of-the-art  in  addressing  aspects  of  the  safety,  reproducibility,  and  efficacy  of  CAP  treatments.  This  dissertation  addresses  two  open  aspects  of  CAP  control,  specifically  designing  feasible  embedded  control  systems  for  point-of-care  CAPJs  and  designing  individualized  CAP  treatment  regimens.  Together,  these  two  aspects  represent  the  overarching  objective  of  this  dissertation:  enabling  point-of-care  devices  for  precision  plasma  medicine.CAPJs  for  biomedical  applications  are  often  touted  for  their  portability  and  point-of-care  use.  Additionally,  medicine  as  a  field  is  moving  towards  more  targeted  approaches  to  patient  healthcare  due  to  the  influx  of  data  from  personal  devices  (e.g.,  smart  wearables)  that  track  health  trends  and  physical  activity  and  due  to  the  importance  of  considering  diverse  patient  profiles  for  equitable  and  efficacious  medical  treatments.  This  trend  (part  of  a  tendency  towards  "edge  computing")  combined  with  the  nonlinear,  multi-variable  CAP  dynamics  calls  for  embedded  control  policies  that  are  capable  of  implementation  on  resource-limited  hardware.  The  first  part  of  this  dissertation  provides  a  novel  fusion  of  hardware  and  software  design  (aka  "hardware-software  co-design")  of  control  policies  to  find  optimal  and  feasible  embedded  control  policies  on  resource-limited  hardware.  In  particular,  key  elements  of  the  end-to-end  design  pipeline  include  the  digital  control  policy,  the  physical  computing  hardware,  and  the  closed-loop  performance  measures  of  interest  such  as  chemical/biological  effects  of  CAPs  on  target  interfaces.  We  demonstrate  that  a  data-driven  optimization  framework  based  on  Bayesian  optimization  (BO),  which  can  simultaneously  incorporate  the  control  policy  design  and  hardware  considerations  when  implementing  the  control  policy,  can  effectively  design  feasible  embedded  control  policies  that  target  multiple  objectives.  An  estimation  of  the  Pareto  frontier  (i.e.,  trade-off  curve)  can  be  generated  via  hardware-in-the-loop  simulations  and  used  to  inform  the  design  of  real-time  control  policies.Several  applications  in  plasma  medicine  require  repeated  treatments  to  realize  therapeutically  effective  treatment  outcomes  to  avoid  overdosing  and/or  to  treat  long-term  conditions.  Prior  works  have  illustrated  predictive  control  strategies  are  capable  of  safely  delivering  CAP  treatments  to  patients,  but  these  strategies  generally  rely  on  underlying  assumptions  of  individual  subject  characteristics  (i.e.,  empirical  models  based  on  population  data).  This  consideration  necessitates  adaptive  treatments  that  are  updated  via  observations  of  treatment  outcomes,  which  can  be  addressed  through  data-driven  optimization.  In  simulations  and  experiments,  we  demonstrated  that  deep  learning-based  control  policies,  which  are  amenable  to  resource-limited  hardware,  can  be  updated  directly  using  multi-objective  BO.  We  demonstrated  how  deep  learning-based  control  policies  can  be  updated  to  find  the  optimal  trade-offs  in  treatment  objectives  when  characteristics  of  individual  subjects  may  differ  from  the  population.  In  a  complementary  direction,  we  developed  a  novel  strategy  to  safely  explore  the  individualized  objective  space  without  compromising  on  performance  improvements.  We  demonstrated  that  our  safe  explorative  BO  strategy  finds  a  balance  between  overly-cautious  exploration  that  may  get  stuck  at  local  optima  and  overly-eager  exploration  that  may  violate  safety-critical  constraints.The  primary  focus  of  this  dissertation  was  on  the  therapeutic  benefits  of  CAPs.  The  final  contribution  of  this  dissertation  investigated  a  novel  aspect  of  CAPs  for  biomedical  use:  (biological)  material  characterization.  We  demonstrated  that  CAPs  are  uniquely  capable  of  producing  minimally  destructive  effects  during  interactions  with  biological  tissues  that  can  be  used  to  identify  and  classify  different  tissue  types.  A  key  aspect  of  this  finding  is  that  real-time  chemical  and  electrical  measurements  of  plasma-tissue  interactions  can  be  analyzed  in  physics-informed  ways  and  fed  into  machine  learning  strategies  to  predict  the  type  of  a  biological  tissue.  Results  from  this  study  can  have  significant  implications  in  non-invasive  early  skin  cancer  detection  systems  and/or  in  real-time  surgical  assistance.To  conclude,  this  dissertation  presented  results  that  illustrate  an  end-to-end  journey  from  the  design  of  physical  computing  hardware  to  the  design  of  digital  control  policies  to  the  design  and  characterization  of  (bio)chemical  outcomes  of  plasma  treatments  in  medicine.  This  dissertation  established  that  data-driven  optimization  is  a  versatile  tool  to  regulate  and  personalize  the  outcomes  of  CAP  treatments.  For  medicine,  BO  mimics  the  doctor-patient  interaction,  and  thus  provides  a  natural  augmentation  to  the  medical  toolkit.  Future  work  may  involve  addressing  additional  challenges  regarding  connected  devices  and  data-driven  strategies  (i.e.,  (cyber)security,  privacy,  distributed  deployment),  fusion  of  physics-structured  learning  with  data,  and  evaluation  of  such  methods  in  preclinical  and  clinical  studies.  The  findings  in  this  dissertation  were  grounded  in  plasma  medicine,  but  can  be  broadly  applicable  to  other  non-equilibrium  plasma  applications,  e.g.,  semiconductor  processing.
■590    ▼aSchool  code:  0028.
■650  4▼aChemical  engineering.
■650  4▼aEngineering.
■650  4▼aInformation  technology.
■650  4▼aComputational  chemistry.
■653    ▼aCold  atmospheric  plasmas
■653    ▼aData-driven  optimization
■653    ▼aMachine  learning
■653    ▼aPlasma  medicine
■653    ▼aPredictive  control
■690    ▼a0542
■690    ▼a0489
■690    ▼a0800
■690    ▼a0537
■690    ▼a0219
■71020▼aUniversity  of  California,  Berkeley▼bChemical  Engineering.
■7730  ▼tDissertations  Abstracts  International▼g86-03B.
■790    ▼a0028
■791    ▼aPh.D.
■792    ▼a2024
■793    ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17161775▼nKERIS▼z이  자료의  원문은  한국교육학술정보원에서  제공합니다.

미리보기

내보내기

chatGPT토론

Ai 추천 관련 도서


    New Books MORE
    Related books MORE
    최근 3년간 통계입니다.

    Detail Info.

    • Reservation
    • 캠퍼스간 도서대출
    • 서가에 없는 책 신고
    • My Folder
    Material
    Reg No. Call No. Location Status Lend Info
    TQ0033694 T   원문자료 열람가능/출력가능 열람가능/출력가능
    마이폴더 부재도서신고

    * Reservations are available in the borrowing book. To make reservations, Please click the reservation button

    해당 도서를 다른 이용자가 함께 대출한 도서

    Related books

    Related Popular Books

    도서위치