org.leialearns.package.html Maven / Gradle / Ivy

Go to download
Show more of this group Show more artifacts with this name
Show all versions of leia-executable Show documentation
The newest version!



    Package 'org.leialearns'



The LEIA system is a learning system that learns from interacting with a completely unknown
    environment over a bidirectional steam of symbols.

The system consists of a few parts that update various models in parallel.


    An {@link org.leialearns.command.encounter.Encounter} communicates directly with the environment and updates
        counters.
    An {@link org.leialearns.command.observer.Observer} consolidates counters into a model of observed behavior.
        The model of observed behavior includes a structure of extra accounting information named deltas
        that reduce the computation effort of other steps.
    A {@link org.leialearns.command.minimizer.Minimizer} applies the
        Minimum Description Length Principle (MDL) to the model of observed behavior, to produce updates
        for the model of expected behavior.
    A {@link org.leialearns.command.consolidator.Consolidator} combines incremental improvements found by the minimizer
        into a consolidated model of expected behavior.
    A maximizer searches for ways to improve the strategy.
    The {@link org.leialearns.command.encounter.Encounter} uses the last strategy to interact with the environment.


The simple, but profound, basis for LEIA is the balance between the designed function of
    the minimizer and the designed function of the maximizer. The minimizer keeps the total
    description length of the model of expected behavior as short as possible while still encoding all
    statistically relevant information about the model of observed behavior. The maximizer, at the other hand,
    searches for strategies that make the total description length of the model of expected behavior
    longer. The maximizer can only increase the length of the model of expected behavior by
    selecting strategies that maximize the expected
    model growth
    given the differences between the model of observed behavior and the model of expected behavior