Algorithmic learning

The growing cell structure algorithm is a variant of a Kohonen network, so the GCS displays several similarities with the SOM. The most distinctive feature of the GCS is that the topology is self-adaptive, adjusting as the algorithm learns about classes in the data. So, unlike the SOM, in which the layout of nodes is regular and predefined, the GCS is not constrained in advance to a particular size of network or a certain lattice geometry. 

Not all computer scientists would agree with this broad statement however, it does encompass virtually every AI method of interest to the chemist. As we shall see, learning is a key part of the definition. In each method discussed in this chapter, there is some means by which the algorithm learns, and then stores, knowledge as it attempts to solve a problem. 

EAs repeatedly carve up old members of the population to create fresh solutions. As in natural selection, competition within the population is essential, otherwise its evolution would be unpredictable and undirected, the algorithm would be as likely to retain poor solutions as promising ones and would make a lengthy and probably unproductive random walk over the search surface. Since individuals in the current population have evolved from those created in past generations, they reflect some of the lessons learned during previous attempts at solution. It is in this fashion that the algorithm learns about a problem. 

Reinforcement learning algorithms Learning algorithms that utilize a system performance measure (that may or may not have a direct, known relationship to output error of the neural network) as a training signal for the neural network. 

Supervisedleaming algorithms Learning algorithms often used in neural networks that use the output error of the neural network as a training signal. 

Definition 3-7 Algorithmic learning is performed by the sole execution of a learning algorithm. Heuristic learning is at least partially based on heuristics. 

K. P. Jantke. Algorithmic learning from incomplete information Principles and problems. In J. Dassow and J. Kelemen (eds). Machines, Languages, and Complexity, pp. 188-207. LNCS 381, Springer-Verlag, 1989. 

S. Muggleton and C. Feng. Efficient induction of logic programs. In Proc. of the 1990 Inf I Workshop on Algorithmic Learning Theory. Ohmsha, 1990. Also in [Muggleton 92], pp. 281-298. 

Identification to finer taxonomic levels will utilize both two-dimensional image information and three-dimensional reconstructions of the specimens. Two-dimensional methods operate directly on one or more images of the specimen, typically taken from preferred views. Our two-dimensional approach is to extract various interest regions and construct invariant descriptors for each region. The descriptors are then clustered and a training algorithm learns feature-cluster associations. Details of two-dimensional classification are provided in Classification Methods (below). 

Figure 1 A schematic of an algorithmic learning approach for microworld control realization. The algorithm is initiated by an optimal design estimate o(t) of the control field, followed by its laboratory refinement in a sequence of experiments coupled to a pattern-recognizing learning algorithm...

---