# US 2003/0225464 A1 · Behavior control apparatus and method

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2003/0225464 A1
- Document ID: US20030225464A1
- Status: Published application
- Kind: A1
- Application number: 10/411,485
- Filed: 2003-04-07
- Published: 2003-12-04
- Earliest claimed: 2002-04-08
- Assignees: Honda Giken Kogyo Kabushiki Kaisha
- Inventors: Yugo Ueda; Hiroshi Tsujino
- Web page: https://patentgrep.com/patent/US20030225464A1

## CPC

- G06N 3/06: Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons
- G06N 3/088: Non-supervised learning, e.g. competitive learning
- G06N 3/0499: Feedforward networks
- G06N 3/08: Learning methods
- G06N 3/09: Supervised learning
- G06N 3/04: Architecture, e.g. interconnection topology
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS

## Abstract

It is objective of the invention to provide a highly reliable control apparatus and method which reduces the amount of calculation required for both learning of an input/output relationship and actual control as well as prevent inappropriate outputs from being generated for inputs which have been never learned. According to one aspect of the invention, pairs of an input pattern vector for learning and a target output are distributed to a class based on the target output. Then, a correspondence between each element of the input pattern vector for learning and the target output is learned only in that class, and a distribution function is calculated for distributing a new input pattern vector for learning to that class. After the completion of this learning, the distribution function is used to determine which class a new input pattern vector detected by a sensor belongs to. Finally, an output is calculated according to the learning result of that class. Therefore, since the range of the outputs corresponding to the inputs may be limited, reliability of the control is improved.

## Claims

### Claim 1

A behavior control apparatus, comprising:
an input unit for creating pairs of an input pattern vector for learning and a target output, wherein said input pattern vector for learning includes inputs detected by a sensor and outputs provided to a controlled object as its elements;
a first distributing unit for distributing said pair of the input pattern vector for learning and the target output to one of two or more classes based on the target output;
a learning unit for learning correspondence between each element in the input pattern vector for learning and the target output in the class where said pair of the input pattern vector and the target output are distributed by said first distributing unit;
a second distributing unit for distributing new input pattern vector into one of said classes according to learning result in said learning unit; and
an output processing unit for calculating an output corresponding to said new input pattern vector according to said learning result in the class where said new input pattern vector is distributed by said second distributing unit, wherein the output is provided to said controlled object.

### Claim 2 (depends on 1)

A behavior control apparatus according to claim 1, wherein said two or more classes are created beforehand.

### Claim 3 (depends on 1)

A behavior control apparatus according to claim 1, wherein said two or more classes are created automatically by self-organizing map method based on a density of distribution of said target output.

### Claim 4 (depends on 1)

A behavior control apparatus according to claim 1, wherein said input pattern vector for learning includes as elements an input detected by the sensor and an output provided to the controlled object at a present time, and an input detected by the sensor and an output to the controlled object at a time preceding the present time.

### Claim 5 (depends on 1)

A behavior control apparatus according to claim 1, wherein:
said learning unit stores each element of said input pattern vectors for learning and said target outputs distributed to each class respectively, and then calculates mean and variance of normal distribution for each data set of said each element and said target output; and
said second distributing unit calculates a distance from the mean of each normal distribution to each element of said new input pattern vector for all classes, and distributes said new input pattern vector to the class which yields smallest mean of the distances.

### Claim 6 (depends on 5)

A behavior control apparatus according to claim 5, wherein said mean of the distances is a weighted mean using weights for each element of said input pattern vector for learning with respect to said target output.

### Claim 7 (depends on 6)

A behavior control apparatus according to claim 6, wherein said weight is calculated based on learning with a neural network with said target output as a teacher signal.

### Claim 8 (depends on 1)

A behavior control apparatus according to claim 1, wherein said output processing unit calculates a weighted mean of distances between mean of each normal distribution in the class where said new input pattern vector is distributed and each element of said new test pattern vector, said output processing unit outputting a value at a distance corresponding to said weighted mean separated from said mean of the normal distribution of said target output.

### Claim 9 (depends on 5)

A behavior control apparatus according to claim 5, wherein:
when each element of said new input pattern vector is represented as xkj (k=1, . . . , K, j=1, . . . , J), said target output is represented as yk, said output corresponding to said new input pattern vector is represented as yk′, said class is represented as i (i=1 to I), said mean and said variance of normal distribution of data set in class i are respectively represented as μij and σij, said weight is represented as wij, and said mean and said variance of normal distribution of said target output yk in class i are represented as μyi, σyi,
said second distributing unit calculates a distance zij from said mean of normal distribution for each element xkj of said new input pattern vector according to following equation:
zij=(xkj−μij)/σij(j=1, . . . , J),
said second distributing unit calculating a weighted mean mi of said calculated distances zij according to following equation:
mi=Σ(zijwij)/J(j=1, . . . , J),
said second distributing unit distributing said new input pattern vector to the class which yields smallest weighted mean mi; and
said output processing unit calculates said output yk′ for the class where said new input pattern vector is distributed according to following equation:
yk=′mi·σyi+μyi.

### Claim 10

A behavior control method, comprising:
creating pairs of an input pattern vector for learning and a target output, wherein said input pattern vector for learning includes inputs detected by a sensor and outputs provided to a controlled object as its elements;
distributing said pair of the input pattern vector for learning and the target output to one of two or more classes based on the target output;
learning correspondence between each element in the input pattern vector for learning and the target output in the class where said pair of the input pattern vector and the target output are distributed;
distributing new input pattern vector into one of said classes according to learning result in said learning step; and
calculating an output corresponding to said new input pattern vector according to said learning result in the class where said new input pattern vector is distributed, wherein the output is provided to said controlled object.

### Claim 11 (depends on 10)

A behavior control method according to claim 10, wherein said two or more classes are created beforehand.

### Claim 12 (depends on 10)

A behavior control method according to claim 10, wherein said two or more classes are created automatically by a self-organizing map method based on a density of distribution of said target output.

### Claim 13 (depends on 10)

A behavior control method according to claim 10, wherein said input pattern vector for learning includes as elements an input detected by the sensor and an output provided to the controlled object at a present time, and an input detected by the sensor and an output to the controlled object at a time preceding the present time.

### Claim 14 (depends on 10)

A behavior control method according to claim 10, wherein:
said step of learning includes storing each element of said input pattern vectors for learning and the target outputs distributed to each class respectively, and then calculating mean and variance of normal distribution for each data set of said each element and said target output; and
said step of distributing said new input pattern vector includes calculating a distance from the mean of each normal distribution to each element of said new input pattern vector for all classes, and distributing said new input pattern vector to the class which yields smallest mean of the distances.

### Claim 15 (depends on 14)

A behavior control method according to claim 14, wherein said mean of the distances is a weighted mean using weights for each element of said input pattern vector for learning with respect to said target output.

### Claim 16 (depends on 15)

A behavior control method according to claim 15, wherein said weight is calculated based on learning with a neural network with said target output as a teacher signal.

### Claim 17 (depends on 10)

A behavior control method according to claim 10, wherein said step of outputting includes calculating a weighted mean of distances between mean of each normal distribution in the class where said new input pattern vector is distributed and each element of said new test pattern vector, and outputting a value at a distance corresponding to said weighted mean separated from said mean of the normal distribution of said target output.

### Claim 18 (depends on 14)

A behavior control method according to claim 14, wherein:
when each element of said new input pattern vector is represented as xkj (k=1, . . . , K, j=1, . . . , J), said target output is represented as yk, said output corresponding to said new input pattern vector is represented as yk′, said class is represented as i (i=1 to I), said mean and said variance of normal distribution of data set in class i are respectively represented as μij and σij, said weight is represented as wij, and said mean and said variance of normal distribution of said target output yk in class i are represented as μyi, σyi,
said step of distributing said new input pattern vector includes calculating a distance zij from said mean of normal distribution for each element xkj of said new input pattern vector according to following equation:
zij=(xkj−μij/σij(j=1, . . . , J)
said step of distributing said new input pattern vector further including calculating a weighted mean mi of said calculated distances zij according to following equation:
mi=Σ(zijwij/J(j=1, . . . , J), and
said step of distributing said new input pattern vector further including distributing said new input pattern vector to the class which yields smallest weighted mean mi, and
said step of outputting includes calculating said output yk′ for the class where said new input pattern vector is distributed according to following equation:
1 y k ′ = m i · σ y i + μ y i .
