RE: [UAI] Non myopic Test selection/value of information

From: Ole-Christoffer Granmo (olegr@simula.no)
Date: Tue May 28 2002 - 08:27:47 PDT

  • Next message: Thomas Lukasiewicz: "[UAI] Reminder: CFP: URWS II at KI2002: Uncertainty in Robotics and Multi-Agent Systems."

    >I'm interested in methods/heuristics for selecting a group
    >of test variables that have a large/largest value of information.
    >Most approaches are myopic (i.e., select the variable with maximal
    >VOI, test its value, select the next one, etc.). I found only two
    >relevant references: Heckerman et al. 1993 discuss approximate
    >non-myopic VOI computation, and Madigan and Russell 1995
    >discuss test selection strategies. Both say that it is a key
    >capability of an expect system, but surprisintly I could not
    >find any work beyond this. In particular, I'm happy with the
    >Heckerman et al. model in which there is a single binary decision
    >and a single binary chance node affecting the value function.
    >
    >Can anyone recommend additional useful references?
    >
    >Thanks,
    >
    >Ronen

    Some work has been done on heuristics for non myopic hypothesis driven data
    request in the context of parallel feature extraction in real-time video
    analysis. This work may be of relevance. The work will be published at the
    2002 International Conference on Parallel and Distributed Processing
    Techniques and Applications (PDPTA'02), Las Vegas, USA, June 2002. The
    abstract of the paper can be found below. If you are interested I can e-mail
    you the paper.

    Ole-Christoffer Granmo

    *************************************************************************

    Real-time Hypothesis Driven Feature Extraction on Parallel Processing
    Architectures

    Ole-Christoffer Granmo and Finn Verner Jensen

    Feature extraction in content-based indexing of media streams is often
    computational intensive. Typically, a parallel processing architecture is
    necessary for real-time performance when extracting features brute force. On
    the other hand, Bayesian network based systems for hypothesis driven feature
    extraction, which selectively extract relevant features one-by-one, have in
    some cases achieved real-time performance on single processing element
    architectures. In this paper we propose a novel technique which combines the
    above two approaches. Features are selectively extracted in parallizable
    sets, rather than one-by-one. Thereby, the advantages of parallel feature
    extraction can be combined with the advantages of hypothesis driven feature
    extraction. The technique is based on a sequential backward feature set
    search and a correlation based feature set evaluation function. In order to
    reduce the problem of higher-order feature-content/feature-feature
    correlation, causally complexly interacting features are identified through
    Bayesian network d-separation analysis and combined into joint features.
    When used on a moderately complex object-tracking case, the technique is
    able to select parallelizable feature sets real-time in a goal oriented
    fashion, even when some features are pairwise highly correlated and causally
    complexly interacting.



    This archive was generated by hypermail 2b29 : Tue May 28 2002 - 08:32:19 PDT