Show simple item record

dc.contributor.authorFonnesbeck, Christopher James
dc.date.accessioned2014-03-03T21:00:55Z
dc.date.available2014-03-03T21:00:55Z
dc.date.issued2003-08
dc.identifier.otherfonnesbeck_christopher_j_200308_phd
dc.identifier.urihttp://purl.galileo.usg.edu/uga_etd/fonnesbeck_christopher_j_200308_phd
dc.identifier.urihttp://hdl.handle.net/10724/21041
dc.description.abstractAs wildlife management goals become more ambitious, associated technical methods must be made available to meet them. Ecologists, conservation biologists and wildlife managers now recognize the need to manage multiple species simul- taneously, or even entire communities, rather than just individual species. The appropriate scale of management is often large, requiring the spatial subdivision of management units. In addition, more researchers have acknowledged the importance of adaptive approaches to management, and the application of multiple working hypotheses in the resolution of management uncertainty. Such requirements speak to the potential analytic complexity of modern wildlife management problems. Cur- rent methods for deriving optimal wildlife resource exploitation and management policies, such as dynamic programming, can become overwhelmed by scenarios of even moderate complexity. This can result in oversimplifcation of management optimization models in order to derive solutions. An investigation of new methods from other disciplines reveals some promising solutions to the \Curse of Dimensionality" which are complimentary to current adaptive resource management approaches. These reside primarily from the artifcial intelligence and machine learning literature, and are correspondingly well-suited to adaptive problems of high dimensionality. This dissertation introduces two of these methods, in the context of adaptive resource management (ARM). In Chapter 1, I discuss the abstraction of dynamic animal populations as stochastic processes, and the study of their control as Markov decision processes. ARM is identifed as a suitable framework for quantifying uncertainties related to management decision- making. Chapter 2 provides an overview of the major forms of constraints that may be applied to the optimization of natural resource management, using the harvest of American black ducks (Anas rubripes) as a case study. Application of optimiza- tion constraints is an important means of satisfying management objectives within ARM. I illustrate the importance of analyzing the effect of imposing constraints on derived optimal control strategies. Unconstrained optimization often yields undesir- able policies, thus revealing implicit objectives; these objectives may be made explicit using constraints. Reinforcement learning (RL) using temporal difference methods is introduced in Chapter 3. This approach to dynamic optimization relies on stochastic approximation to derive policies that are asymptotically optimal as they are updated with either simulated or real experience with the system of interest. Importantly, applications of RL to machine learning problems have shown them capable of opti- mizing very complex decision-making problems, with large numbers of decisions, states and stochastic variables. I demonstrate RL for the first time in the context of natural resource management using the famous mallard (Anas platyrhynchos) har- vest management case study by Anderson (1975). Though convergence to the true optimal policies (derived from dynamic programming) did not occur, the estimated policies achieved nearly identical objective return relative to the optimal policies. Chapter 4 reviews Kalman filtering methods, and their applicability to ARM through the reduction of model parameter uncertainty. In particular, the recent integration of unscented transformations for recursively updating the parameters of non-linear systems makes Kalman filtering now more widely applicable to ecological problems. These are just two of a growing number of modern methods for the identification and control of non-linear dynamic systems that exist across the disciplines of arti- ficial intelligence, engineering and computer science. Though challenges exist for integrating these new methods into ARM, they show promise as tools for solving more complex optimal wildlife management decision-making problems.
dc.languageeng
dc.publisheruga
dc.rightspublic
dc.subjectAdaptive resource management
dc.subjectconstraints
dc.subjectharvest management
dc.subjectKalman filtering
dc.subjectmachine learning
dc.subjectoptimization
dc.subjectreinforcement learning
dc.subjectstochastic approximation
dc.subjectwildlife management
dc.subjectwaterfowl
dc.titleAdaptive management of dynamic wildlife resource systems
dc.title.alternativeoptimization, constraints and learning
dc.typeDissertation
dc.description.degreePhD
dc.description.departmentForest Resources
dc.description.majorForest Resources
dc.description.advisorMichael J. Conroy
dc.description.committeeMichael J. Conroy
dc.description.committeeJohn P. Carroll
dc.description.committeeDANIEL B. HALL
dc.description.committeeJames T. Peterson
dc.description.committeeJaxk H. Reeves


Files in this item

FilesSizeFormatView

There are no files associated with this item.

This item appears in the following Collection(s)

Show simple item record