Learning apparatus and method

• US 20070112708A1
• Filed: 10/30/2006
• Published: 05/17/2007
• Est. Priority Date: 10/31/2005
• Status: Active Grant

• Current Assignee: Sony Corporation (Sony Group Corp.)
• Original Assignee: Sony Corporation (Sony Group Corp.)
• Inventors: Sawada, Tsutomu
• Granted Patent: US 7,720,774 B2
• Time in Patent Office: Days
• US Class Current: 706/25
• CPC Class Codes: G06N 7/01 Probabilistic graphical models, e.g. probabilistic networks

Abstract

A learning apparatus for building a network structure of a Bayesian network based on learning data. In the Bayesian network, a cause and effect relationship between plural nodes is represented by a directed graph. The learning apparatus includes a storage portion in which the learning data is stored and a learning portion for building the network structure based on the learning data. The learning portion prepares an initial population of individuals formed by individuals each having a genotype in which orders between the nodes and cause and effect relationship have been stipulated, repeatedly performs processing for crossovers and/or mutations on the initial population of individuals based on a genetic algorithm, calculates an evaluated value of each individual based on the learning data, searches for an optimum one of the individuals, and takes a phenotype of the optimum individual as the network structure.

Claims

1. A learning apparatus for building, based on learning data, a network structure of a Bayesian network in which a cause and effect relationship between plural nodes is represented by a directed graph, the learning apparatus comprising:

storage means in which the learning data is stored; and

learning means for building the network structure based on the learning data;

wherein the learning means prepares an initial population of individuals constituted by individuals each having a genotype in which orders between the nodes and cause and effect relationship have been stipulated, repeatedly performs processing for crossovers and/or mutations on the initial population of individuals based on a genetic algorithm, calculates an evaluated value of each of the individuals based on the learning data, searches for an optimum one of the individuals, and takes a phenotype of the optimum individual as the network structure.

2. A learning apparatus as set forth in claim 1, wherein the genotype takes the plural nodes arranged in a first direction according to a stipulated order as parent nodes, takes the plural nodes arranged in a second direction perpendicular to the first direction according to the stipulated order as child nodes, and stipulates presence or absence of a cause and effect relationship between corresponding nodes by alleles in each gene locus at which each parent node and each child node correspond to each other.

3. A learning apparatus as set forth in claim 2, wherein genes at gene loci at which child nodes have orders equal to or higher than orders of parent nodes show no manifestation of traits.

4. A learning apparatus as set forth in claim 2, wherein in a case where the number of parent nodes having a cause and effect relationship with an arbitrary child node is excess of a give number, the learning means inverts genes in at least one gene locus associated with the child node onto its allele such that the number of parent nodes having the cause and effect relationship with the child node becomes less than the given number.

5. A learning apparatus as set forth in claim 1, wherein the initial population of individuals reflects all or some of prior knowledge of a designer.

6. A learning apparatus as set forth in claim 1, wherein the initial population of individuals is based on individuals obtained as a result of learning.

7. A learning apparatus as set forth in claim 1, wherein the evaluated value is a logarithm of a BD metric, and wherein when an ith node of n nodes is taken as a child node Xi, the child node X_i can assume r_i values from v_io to v_iri-1, and the number of patterns of values that all parent nodes having a cause and effect relationship with the child node X_i can assume is q_i, if the value of the child node X_i is v_ik and data is produced in which values of all the parent nodes having the cause and effect relationship with the child node X_i assume a jth pattern, a number of times N′

_ijk which is anticipated prior to acquisition of the learning data is calculated according to the following formula;

$N_{\mathrm{ijk}}^{\mathrm{\prime <br><br>}}=\left(\underset{i=0}{\overset{n-1}{\prod <br><br>}}<br><br>\text{\hspace{1em}<br><br>}<br><br>r_i\right)/r_i<br><br>q_i$

8. A learning apparatus as set forth in claim 1, wherein the evaluated value is a logarithm of a BD metric, and wherein when an ith node of n nodes is taken as a child node X_i, the child node X_i can assume r_i values from v_io to v_iri-1, and the number of patterns of values that all parent nodes having a cause and effect relationship with the child node X_i can assume is q_i, if the value of the child node X_i is v_ik and data is produced in which values of all the parent nodes having the cause and effect relationship with the child node X_i assume a jth pattern, a number of times N′

_ijk which is anticipated prior to acquisition of the learning data is set equal to zero (N′

_ijk=0).

9. A learning method of building, based on learning data, a network structure of a Bayesian network in which a cause and effect relationship between plural nodes is represented by a directed graph, the learning method comprising the steps of:

preparing an initial population of individuals constituted by individuals each having a genotype in which orders between the plural nodes and cause and effect relationship have been stipulated;

repeatedly performing processing for crossovers and/or mutations on the initial population of individuals based on a genetic algorithm;

calculating an evaluated value of each of the individuals based on the learning data;

searching for an optimum one of the individuals; and

taking a phenotype of the optimum individual as the network structure.

10. A learning apparatus for building, based on learning data, a network structure of a Bayesian network in which a cause and effect relationship between plural nodes is represented by a directed graph, the learning apparatus comprising:

a storage portion in which the learning data is stored; and

a learning portion operable to build the network structure based on the learning data;

wherein the learning portion prepares an initial population of individuals constituted by individuals each having a genotype in which orders between the nodes and cause and effect relationship have been stipulated, repeatedly performs processing for crossovers and/or mutations on the initial population of individuals based on a genetic algorithm, calculates an evaluated value of each of the individuals based on the learning data, searches for an optimum one of the individuals, and takes a phenotype of the optimum individual as the network structure.