ALGORITHM (Iterative refinement) 5)
← Back
An algorithm which progressively produce better answers when iterated.
As "the backpropagation algorithm is just a steepest descent approach without any line search… it inherits the well known disadvantages of the gradient method: The convergence depends strongly on the starting values of the parameters. The direction of the local steepest descent is often far different from the overall direction where the minimum lies" and… "The algorithm… normally either overshoots the minimum (with the risk of oscillations) or – if the gradient is small – there is hardly any progress at all" (TRONCALE, 1985, p.449).
This difficulty is alleviated by the introduction of iterative refinement algorithms.
Such algorithms are constructed from a set of rules whose operation is interactive and not rigorously preprogramed.
It is however open to debate if such a kind of mental device really fits within the concept of algorithm.
Categories
- 1) General information
- 2) Methodology or model
- 3) Epistemology, ontology and semantics
- 4) Human sciences
- 5) Discipline oriented
Publisher
Bertalanffy Center for the Study of Systems Science(2020).
To cite this page, please use the following information:
Bertalanffy Center for the Study of Systems Science (2020). Title of the entry. In Charles François (Ed.), International Encyclopedia of Systems and Cybernetics (2). Retrieved from www.systemspedia.org/[full/url]
We thank the following partners for making the open access of this volume possible: