Repository logo

Browsing by Author "Whitley, L. Darrell, committee member"

Now showing 1 - 3 of 3
  • Thumbnail Image
    ItemOpen Access
    A synthesis of reinforcement learning and robust control theory
    (Colorado State University. Libraries, 2000) Kretchmar, R. Matthew, author; Anderson, Charles, advisor; Howe, Adele E., committee member; Whitley, L. Darrell, committee member; Young, Peter M., committee member; Hittle, Douglas C., committee member
    The pursuit of control algorithms with improved performance drives the entire control research community as well as large parts of the mathematics, engineering, and artificial intelligence research communities. A fundamental limitation on achieving control performance is the conflicting requirement of maintaining system stability. In general, the more aggressive is the controller, the better the control performance but also the closer to system instability. Robust control is a collection of theories, techniques, the tools that form one of the leading edge approaches to control. Most controllers are designed not on the physical plant to be controlled, but on a mathematical model of the plant; hence, these controllers often do not perform well on the physical plant and are sometimes unstable. Robust control overcomes this problem by adding uncertainty to the mathematical model. The result is a more general, less aggressive controller which performs well on the both the model and the physical plant. However, the robust control method also sacrifices some control performance in order to achieve its guarantees of stability. Reinforcement learning based neural networks offer some distinct advantages for improving control performance. Their nonlinearity enables the neural network to implement a wider range of control functions, and their adaptability permits them to improve control performance via on-line, trial-and-error learning. However, neuro-control is typically plagued by a lack of stability guarantees. Even momentary instability cannot be tolerated in most physical plants, and thus, the threat of instability prohibits the application of neuro-control in many situations. In this dissertation, we develop a stable neuro-control scheme by synthesizing the two fields of reinforcement learning and robust control theory. We provide a learning system with many of the advantages of neuro-control. Using functional uncertainty to represent the nonlinear and time-varying components of the neuro networks, we apply the robust control techniques to guarantee the stability of our neuro-controller. Our scheme provides stable control not only for a specific fixed-weight, neural network, but also for a neuro-controller in which the weights are changing during learning. Furthermore, we apply our stable neuro-controller to several control tasks to demonstrate that the theoretical stability guarantee is readily applicable to real-life control situations. We also discuss several problems we encounter and identify potential avenues of future research.
  • Thumbnail Image
    ItemOpen Access
    Automatic endpoint vulnerability detection of Linux and open source using the National Vulnerability Database
    (Colorado State University. Libraries, 2008) Whyman, Paul Arthur, author; Ray, Indrajit, advisor; Krawetz, Neal, committee member; Whitley, L. Darrell, committee member; Hayne, Stephen, committee member
    A means to reduce security risks to a network of computers is to manage which computers can participate on a network, and control the participation of systems that do not conform to the security policy. Requiring systems to demonstrate their compliance to the policy can limit the risk of allowing uncompiling systems access to trusted networks. One aspect of determining the risk a system represents is patch-level, a comparison between the availability of vendor security patches and their application on a system. A fully updated system has all available patches applied. Using patch level as a security policy metric, systems can evaluate as compliant, yet may still contain known vulnerabilities, representing real risks of exploitation. An alternative approach is a direct comparison of system software to public vulnerability reports contained in the National Vulnerability Database (NVD). This approach may produce a more accurate assessment of system risk for several reasons including removing the delay caused by vendor patch development and by analyzing system risk using vender-independent vulnerability information. This work demonstrates empirically that current, fully patched systems contain numerous software vulnerabilities. This technique can apply to platforms other than those of Open Source origin. This alternative method, which compares system software components to lists of known software vulnerabilities, must reliably match system components to those listed as vulnerable. This match requires a precise identification of both the vulnerability and the software that the vulnerability affects. In the process of this analysis, significant issues arose within the NVD pertaining to the presentation of Open Source vulnerability information. Direct matching is not possible using the current information in the NVD. Furthermore, these issues support the belief that the NVD is not an accurate data source for popular statistical comparisons between closed and open source software.
  • Thumbnail Image
    ItemOpen Access
    Geometric methods on special manifolds for visual recognition
    (Colorado State University. Libraries, 2010) Lui, Yui Man, author; Beveridge, J. Ross, advisor; Kirby, Michael, 1961-, committee member; Draper, Bruce A. (Bruce Austin), 1962-, committee member; Whitley, L. Darrell, committee member
    Many computer vision methods assume that the underlying geometry of images is Euclidean. This assumption is generally not valid. Therefore, this dissertation introduces new nonlinear geometric frameworks based upon special manifolds, namely Graβmann and Stiefel manifolds, for visual recognition. The motivation for this thesis is driven by the intrinsic geometry of visual data in which the visual data can be either a still image or video. Visual data are represented as points in appropriately chosen parameter spaces. The idiosyncratic aspects of the data in these spaces are then exploited for pattern classification. Three major research results are presented in this dissertation: face recognition for illumination spaces on Stiefel manifolds, face recognition on Graβmann registration manifolds, and action classification on product manifolds. Previous work has shown that illumination cones are idiosyncratic for face recognition in illumination spaces. However, it has not been addressed how a single image relates to an illumination cone. In this dissertation, a Bayesian model is employed to relight a single image to a set of illuminated variants. The subspace formed by these illuminated variants is characterized on a Stiefel manifold. A new distance measure called Canonical Stiefel Quotient (CSQ) is introduced. CSQ performs two projections on a tangent space of a Stiefel manifold and uses the quotient for classification. The proposed method demonstrates that illumination cones can be synthesized by relighting a single image to a set of images, and the synthesized illumination cones are discriminative for face recognition. Experiments on the CMU-PIE and YaleB data sets reveal that CSQ not only achieves high recognition accuracies for generic faces but also is robust to the choice of training sets. Subspaces can be realized as points on Graβmann manifolds. Motivated by image perturbation and the geometry of Graβmann manifolds, we present a method called Graβmann Registration Manifolds (GRM) for face recognition. First, a tangent space is formed by a set of affine perturbed images where the tangent space admits a vector space structure. Second, the tangent spaces are embedded on a Graβmann manifold and chordal distance is used to compare subspaces. Experiments on the FERET database suggest that the proposed method yields excellent results using both holistic and local features. Specifically, on the FERET Dup2 data set, which is generally considered the most difficult data set on FERET, the proposed method achieves the highest rank one identification rate among all non-trained methods currently in the literature. Human actions compose a series of movements and can be described by a sequence of video frames. Since videos are multidimensional data, data tensors are the natural choice for data representation. In this dissertation, a data tensor is expressed as a point on a product manifold and classification is performed on this product space. First, we factorize a data tensor using a modified High Order Singular Value Decomposition (HOSVD) and recognize each factorized space as a Graβmann manifold. Consequently, a data tensor is mapped to a point on a product manifold and the geodesic distance on the product manifold is computed for tensor classification. The proposed method is geometrically sound and the metric is naturally inherited from the factor manifolds. Experiments on the Cambridge-Gesture and KTH human action data sets show that the proposed method outperforms the current state-of-the-art. The use of special manifolds for visual recognition has just emerged. This dissertation shows that the underlying geometry of space is an important feature for pattern recognition. The proposed geometric frameworks are particularly suitable for high dimensional data, and will lead to many possible future work.