JP3848076B2 - Virtual biological system and pattern learning method in virtual biological system - Google Patents

Description

[0001]
BACKGROUND OF THE INVENTION
The present invention relates to a virtual biological system that recognizes an input pattern from the outside, and more particularly to a virtual biological system that spontaneously learns a pattern and a pattern learning method in the virtual biological system.
[0002]
[Prior art]
In recent years, various software applications in which characters imitating organisms appear and robots imitating organisms have been proposed. These software applications and robots, which should be called virtual biological systems, are designed to output responses such as expressing emotions according to input information from the user. Some of these software applications and robots have a capability of recognizing pattern information such as a user's face image and voice waveform, and can respond to such pattern recognition information as input information.
[0003]
For example, see http: // www. incx. nec. co. The “help robot R100” disclosed in jp / robot / is capable of identifying who the person is from the face image of the user and performing an operation specific to the user based on the recognition result. .
[0004]
Also, for example, http: // www. lares. dti. ne. jp / ~ p-chamin / The software application called “Pechamin” took a voice signal from a microphone connected to a personal computer, learned it and listened to it, and processed the voice signal. It is an “Inco” type virtual creature that can utter a waveform signal as its own voice. Furthermore, “Pichamin” learns the context of individual voices that are input separately, and can generate the self-speech in the same context.
[0005]
The basic framework for pattern recognition is to express the characteristics of patterns belonging to the class to be recognized in some form and to verify how well the characteristics of unknown human power patterns match the characteristics of this class. . Information representing the characteristics of this class is called a dictionary, and a scale indicating the degree of matching is called similarity. When there is one class to be recognized (this is called an identification problem), if this similarity is greater than or equal to a predetermined threshold, the unknown input pattern is recognized as belonging to this class.
[0006]
In addition, when there are two or more classes to be recognized (this is called an identification problem), if there is a class whose similarity is equal to or higher than a predetermined threshold and the highest similarity is obtained, the unknown input pattern has the highest similarity. Acknowledge that you belong to the class In any case, if there is no class equal to or greater than the predetermined threshold, the unknown input pattern is determined to be in a class undecided, that is, belongs to a new class that cannot be recognized by the system.
Here, the calculation mechanism of pattern recognition will be briefly described.
[0007]
In general, the pattern p is data composed of N scalar quantities. Therefore, in the pattern p, each scalar quantity corresponds to one axis, and the value of each scalar quantity is a coordinate value on the corresponding axis. In the N-dimensional superspace S <a space having N orthogonal coordinate axes> It can be regarded as a point (or a position vector v from the origin). In the case of an image pattern, each pixel value of an image whose size and brightness are normalized can correspond to this scalar quantity. In the case of an audio pattern, the spectral component at each time with the time length and intensity normalized can be made to correspond to the scalar quantity. This is a pattern vectorization method used in many pattern recognition apparatuses.
[0008]
At this time, whether the two patterns are so similar can be evaluated by calculating the similarity between the patterns. There are various ways of defining the similarity between two patterns. For example, the cosine of a normalized position vector (referred to as a feature vector) representing the length of each pattern (ie, projection length from one to the other) The similarity is defined as In this case, the greater the similarity, the more similar the two patterns.
[0009]
In general, when a plurality of known patterns (called teaching patterns) belonging to a certain class are given and one unknown input pattern is given, the possibility that this unknown input pattern belongs to this class is evaluated. Is simply calculated by calculating the similarity between the two patterns. In this case, the similarity between the unknown input pattern and all the known patterns is obtained individually, and for example, the largest similarity is set as the similarity with this class. In this method, all known patterns are individually stored as a dictionary representing class features.
[0010]
However, with this method, as the number of known patterns increases, the cost of similarity calculation, the memorandum base for storing known patterns, and the like will increase without limit. In addition, since there are patterns similar to each other in the known patterns, there is a waste of calculating the degree of similarity many times with similar patterns. Therefore, instead of storing the known patterns individually, it is necessary to describe the feature of the known pattern set, that is, the class with less information.
[0011]
As a method of expressing the characteristics of a class, K known patterns belonging to a certain class, that is, K N-dimensional feature vectors {v: v1,. . , VK} (corresponding to the above-mentioned known pattern), M (M <N) orthonormal vectors {e: e1,. . . , EM} as a base, there is a method using an M-dimensional superspace L as a dictionary. This M-dimensional superspace L is a subspace L of M dimensions of the N-dimensional superspace S (because M <N), and this M-dimensional superspace L is called a dictionary subspace of M dimensions.
[0012]
According to this method, even if the number K of known patterns is much larger than N, it is only necessary to store at most M (<N) basis vectors as a dictionary. When one unknown feature vector v (corresponding to the above-mentioned unknown pattern) is given, the similarity to the dictionary subspace L is determined by using the unknown feature vector v as the basis vector {e: e1,. . . , EM} is defined as the similarity to the sum of squares of the length projected onto each of e, M}. This is called the subspace method.
[0013]
Detailed information related to the definition of the dictionary expression and similarity including the subspace method can be found in Ref. [1] (Erki Oya, translated by Hidemitsu Ogawa, “Pattern Recognition and Subspace”, Sangyo Tosho, 1986. ), [2] <Yasuzo Iijima, “Pattern Recognition Theory”, Morikita Publishing, 1989).
[0014]
As exemplified above, in a system that generally recognizes some pattern, it is necessary to give the system a dictionary of classes to be recognized. Normally, it is difficult to create this dictionary synthetically regardless of the actual pattern. Therefore, the actual pattern (the above-mentioned teaching pattern) belonging to the target class is collected, and the dictionary is analyzed by principal component analysis in the illustrated subspace method. Will be generated.
[0015]
At this time, in order to generate a dictionary from an actual pattern, at least a teaching pattern for constructing the dictionary and information of a class to which the teaching pattern belongs are necessary. The teaching pattern is the face image pattern for face recognition, the character image pattern for character recognition, the word waveform pattern for word speech recognition, and the class information is the name (character code string) and identification of the person for personal recognition by face. If the facial expression is recognized by a face, the name of the facial expression (character code string) and identification number, the character code of the character is recognized by character recognition, the character code string of the word is generally recognized by word speech recognition, and the like. Are numbers and symbols (columns) that can be individually identified.
[0016]
In this way, a necessary dictionary of a necessary class can be generated only by providing two types of information, that is, a teaching pattern and information on a class to which the teaching pattern belongs.
[0017]
It should be noted that generating such a dictionary, that is, adding a new dictionary or updating an existing dictionary at any time during operation of the pattern recognition system and improving the pattern recognition ability of the system will be referred to as pattern learning hereinafter.
[0018]
Conventionally, it has been necessary for humans to provide two types of information, that is, teaching patterns and class information, to the system. In particular, in a pattern recognition system that can perform pattern learning during operation, a human collects teaching patterns while using the pattern input function for pattern recognition, and manually inputs the class to which the pattern belongs. Pattern learning proceeds in the procedure of creating a dictionary with the help of the system. Therefore, in addition to the original function of inputting and recognizing a pattern in such a system, an operation input function for calling a mode for collecting a teaching pattern as a dictionary generation function, and an input pattern in the mode as a teaching pattern A function of storing, a function of accepting an operation input of class information, and a function of generating a dictionary of a specified class from a collected and stored teaching pattern are provided.
[0019]
As is clear from this, in the conventional pattern learning, the input of the teaching timing of “Learning the pattern to be input from now on”, the pattern presentation to the system for the teaching performed intentionally after that, “Teached Human intervention was indispensable for the three stages of class information input, “Pattern belongs to this class”.
[0020]
However, in this way, it was impossible to produce the virtual creature's creature-like learning ability to learn things without knowing it while using it by learning the pattern manually. This is a major obstacle to providing a more biological artificial system.
[0021]
Also in the above-mentioned “R100”, face teaching must be performed explicitly, and manual teaching work is necessary to learn a new face.
[0022]
In addition, in the above-mentioned “Piichi Chamin”, when teaching the voice that the user wants to learn, the teaching voice is heard after clicking the microphone button displayed on the screen with the mouse. It was necessary to explicitly notify the system of the pattern to be learned. As described above, the conventional pattern recognition system capable of pattern learning requires intentional work performed only for teaching by a human in pattern learning.
[0023]
However, the only exception is “Pechamin” mentioned above, and the system also has a function to autonomously learn the voice picked up by the microphone, enabling pattern learning that does not require explicit work by the user. Yes. When "Pichamin" recognizes a new pattern that is not similar to any of the voice patterns learned so far, it automatically issues new class information (probably a new identification number) to that pattern, thereby providing class information for that pattern. It seems that the dictionary that is generated internally and recognizes this new pattern is automatically generated. With this function, “Piichi Chamin” has succeeded in producing a virtual creature that learns words and sounds, even if the user does nothing.
[0024]
However, even in “Pichamin” with such a spontaneous learning ability, there are insufficient means to determine which speech to be picked up by the microphone. This causes a problem that the user speaks an unknown voice. This means that there is no appropriate means for selecting the pattern to be learned, which is a problem of the prior art.
[0025]
[Problems to be solved by the invention]
The present invention has been made in view of the above-described problems in learning of virtual creatures, and the object of the present invention is to create a virtual creature that can perform spontaneous pattern learning by eliminating inappropriate patterns. It is to provide a pattern learning method in a system and a virtual biological system.
[0026]
[Means for Solving the Problems]
The present invention is characterized in that the input pattern information is recognized and the internal state is updated based on the result.
[0027]
In order to achieve the above object, according to claim 1 of the present invention, input means for inputting pattern information from a plurality of types of signal sources, and the input pattern information. Based on dictionary information Based on the recognition result by the pattern recognition means and the pattern recognition means Emotion State updating means for updating the state; and Emotion Output means for outputting a response according to the state Do In the virtual biological system, Emotion When the condition satisfies a predetermined condition, New dictionary information for recognizing the pattern information, an amount of updating the emotional state by detecting the pattern information, Learning means to learn further Equipped Do A virtual biological system characterized by the above is provided.
[0028]
Therefore, the input pattern is selectively learned according to the condition of the internal state, and a virtual biological system that learns only an appropriate input pattern and performs a reasonable output can be obtained.
[0029]
According to claim 2 of the present invention, input processing for inputting pattern information from a plurality of types of signal sources, and the input pattern information Based on dictionary information Based on the pattern recognition processing to be recognized and the recognition result by the pattern recognition means Emotion State update processing for updating the state, and Emotion Output processing that outputs a response according to the state A pattern learning method in a virtual biological system, comprising: Said Emotion When the condition satisfies a predetermined condition, New dictionary information for recognizing the pattern information, an amount of updating the emotional state by detecting the pattern information, Pattern learning process to learn Further have A pattern learning method in a virtual biological system is provided.
[0030]
The input pattern is selectively learned according to the condition of the internal state, and a rational pattern learning method for learning only an appropriate input pattern is obtained.
[0031]
According to claim 3 of the present invention, in the virtual biological system according to claim 1, The predetermined condition is that the intensity of the emotional state exceeds a predetermined threshold value Provide virtual biological systems.
[0032]
According to claim 4 of the present invention, in the pattern learning method according to claim 2, the predetermined condition is: Emotion Provided is a pattern learning method in a virtual biological system, characterized in that the intensity of a state exceeds a predetermined threshold.
[0033]
According to claim 5 of the present invention, input means for inputting pattern information from one or more signal sources, and pattern information input by the input means are divided into a plurality of groups. For each group based on dictionary information categorized into Pattern recognition means for recognizing New dictionary information for recognizing the pattern information, and an amount of updating the emotional state by detecting the pattern information. Based on the pattern learning means for learning and the recognition result by the pattern recognition means Emotion State update means for updating the state; and Emotion When the state satisfies a predetermined condition, a learning command unit that gives a learning command to the pattern learning unit; Emotion Output means for outputting a response according to the state Do A virtual biological system characterized by the above is provided.
[0034]
According to claim 6 of the present invention, input means for inputting pattern information from one or more signal sources, and pattern information input by the input means are divided into a plurality of groups. For each group based on dictionary information classified into Pattern recognition means for recognizing New dictionary information for recognizing the pattern information, an amount of updating the emotional state by detecting the pattern information, Based on the pattern learning means for learning and the recognition result by the pattern recognition means, it is represented by a plurality of numerical values. Emotion By changing these numbers of states Emotion State update means for updating the state; and Emotion Learning command means for giving a learning command to the pattern learning means when a plurality of numerical values representing a state has a predetermined size; Emotion Provided is a virtual biological system comprising output means for outputting a response according to a state.
[0035]
According to claim 7 of the present invention, input means for inputting pattern information from one or more signal sources, and a plurality of pieces of pattern information input within a predetermined time by the input means are divided into a plurality of groups. For each group based on dictionary information categorized into Pattern recognition means for recognizing New dictionary information for recognizing the pattern information, an amount of updating the emotional state by detecting the pattern information, Based on the pattern learning means for learning and the recognition result by the pattern recognition means, it is represented by a plurality of numerical values. Emotion By changing these numbers of states Emotion State update means for updating the state; and Emotion Learning command means for giving a learning command to the pattern learning means when a plurality of numerical values representing a state has a predetermined size; Emotion Output means for outputting a response according to the state Do A virtual biological system characterized by the above is provided.
[0036]
Here, the “predetermined time period during which a plurality of pieces of pattern information are input by the input means” usually refers to a time period during which the user operates as a series of operations. According to the present invention, by a plurality of pattern information Emotion Since the state is changed, a virtual biological system that operates more rationally can be obtained.
[0037]
According to claim 8 of the present invention, claim 6 Or claims The virtual biological system according to claim 7, wherein the Emotion The virtual biological system is characterized in that the plurality of numerical values representing the state are numerical values representing at least happiness, excitement, and likability.
[0038]
According to claim 9 of the present invention, claim 6 Or claims 8. The virtual biological system according to claim 7, wherein the pattern information is recognized for at least one of a group of people and vocabulary and at least one of a group of facial expressions, vocabulary, and force sense. provide.
[0039]
According to claim 10 of the present invention, claim 6 Or claims 7. The virtual biological system according to claim 7, wherein the pattern information is divided into groups of person, facial expression, vocabulary, vocabulary, and force sense, and among these, the group of facial expression, vocabulary, and force sense about Of stimulation Recognition result Pleasant and uncomfortable, said Emotion Provided is a virtual biological system characterized by changing the numerical values of happiness, excitement, and likability representing a state.
[0040]
According to the eleventh aspect of the present invention, the step of inputting pattern information from one or more signal sources, and the plurality of pattern information input within a predetermined time by this step into a plurality of groups. For each group based on classified dictionary information Recognized by a plurality of numerical values based on the pattern recognition step to be recognized and the recognition result by this pattern recognition step Emotion By changing these numbers of states The emotion A state update step for updating the state; Emotion When at least one of a plurality of numerical values representing a state reaches a predetermined size, New dictionary information for recognizing the pattern information, an amount of updating the emotional state by detecting the pattern information, In accordance with a learning command step for giving a command for learning and a command for this learning command step The new dictionary information and the update amount A pattern learning step for learning Emotion An output step that outputs a response according to the state Have A pattern learning method in a virtual biological system is provided.
[0041]
According to a twelfth aspect of the present invention, in the pattern learning method according to the eleventh aspect, the pattern information is divided into groups of persons, facial expressions, vocabulary, vocabulary, and force senses. And about force groups Stimulating Due to the pleasant and uncomfortable recognition results, Emotion Provided is a pattern learning method in a virtual biological system, characterized in that numerical values of happiness, excitement, and favorableness representing a state are changed.
[0042]
DETAILED DESCRIPTION OF THE INVENTION
Hereinafter, an embodiment of a virtual biological system according to the present invention will be described with reference to the drawings.
[0043]
FIG. 1 is a diagram showing a functional block configuration of a virtual biological system according to an embodiment of the present invention. This virtual biological system includes an input unit 1 that receives information from the outside world, a pattern recognition unit 2 that recognizes the input information as a pattern, a state update unit 3 that updates the internal state of the system, and an external action An output unit 4, a learning command generation unit 5 for instructing learning of the pattern of the information, and a pattern learning unit 6 for performing pattern learning based on this instruction.
[0044]
FIG. 2 is a diagram showing the flow of processing in this virtual biological system. The processing in the virtual biological system of this embodiment includes an input processing step S1, a pattern recognition processing step S2, a state update processing step S3, an output processing step S4, a learning determination processing step S5, and a pattern learning processing step S6.
[0045]
The input unit 1 is disposed on the exterior of the apparatus, such as a post-screen input means such as a TV camera for inputting a user's face image pattern, a microphone or the like for inputting a user's voice waveform pattern, and contact from the user. Force input means such as a pressure sensor for inputting force patterns is provided. Pattern information (referred to as input patterns) from a plurality of types of signal sources such as sound signals and force signals, and sound signals and image signals. Is a means for inputting.
[0046]
FIG. 3 shows a specific arrangement example of various input means of the embodiment when the present invention is applied to a robot 11 simulating an animal. This robot 11 has a television camera 12 as an image input means at a part corresponding to the eyes, a microphone 13 as a voice input means at a part corresponding to the ears, and a force sense input means at a part corresponding to the top of the head. A pressure sensor 14 is provided.
[0047]
The virtual biological system in the present invention is not limited to a robot that specifically exists as shown in FIG. 3, but may be a computer graphic (CG) that appears on a computer screen.
[0048]
FIG. 4 shows an arrangement example of various input means when applied to the display virtual creature 22 by CG displayed on the display device of the computer 21 in this way. In this example, a television camera 23 serving as an image input unit is installed in front of the computer 21, and a microphone 24 serving as a voice input unit is installed beside the computer 21.
[0049]
The force sense input means is the left button 26 of the mouse 25. By reciprocating the mouse cursor 27 on the virtual creature 22 while pressing the left button 26, it is possible to give an input with a scissors meaning an action of stroking the display virtual creature 22. In addition, when the left button 26 is clicked on the display virtual creature 22 for a short time, it is possible to perform a hit input meaning an operation of hitting the display virtual creature 22. Furthermore, by pressing down the left button 26 on the display virtual creature 22, it is possible to perform a pressing input that means an operation of pressing the virtual creature 22.
[0050]
The input process is performed in step S11 of FIG. 2 by the means described above.
[0051]
The pattern recognition unit 2 in FIG. 1 holds dictionary information for pattern recognition, collates the input pattern by the input unit 1 with the dictionary information of each class using a subspace method, and has the highest similarity above a predetermined threshold. 2 is a means for detecting a class indicating a degree, certifying the class as a class to which the input pattern belongs, and outputting identification information of the certified class. In step S12 of FIG. 2, such pattern recognition processing is performed. The
[0052]
When the pattern recognition unit 2 cannot find a class to be recognized, the pattern recognition unit 2 outputs special identification information indicating that the input pattern is an unlearned class.
[0053]
The pattern recognition unit 2 touches the user's face image pattern according to the person and the facial expression, the user's voice spectrum pattern according to the vocabulary and the vocabulary according to the user's contact pressure pattern. You can recognize another. Therefore, as shown in FIG. 5, the pattern recognition unit 2 can hold dictionary information classified into five types of groups of people, facial expressions, vocabulary, vocabulary, and force sense, and outputs recognition results for each group. it can.
[0054]
In FIG. 5, the initial number of dictionaries is the number of dictionaries from the beginning. If this number is 0, it means that there is no dictionary in the group at first.
[0055]
For each group of facial expressions, vocabulary, and force sense, the initial number of dictionaries is 2 or 3, and corresponding dictionary information is prepared in advance so that the system can naturally recognize them. On the other hand, for the group of people and vocabulary, since the initial dictionary number is 0, dictionary information is not prepared in advance so that its recognition ability should be acquired during operation, and it remains innocent. At this time, a class that can be recognized naturally is called an initial class.
[0056]
In this embodiment, the format of the identification information of each group is defined as shown in FIG. 6 as the identification information obtained as the recognition result. That is, the identification information of each group of person, facial expression, vocabulary, vocabulary and force sense is represented by adding a serial number starting from 1 after the FI, FE, VI, VE, and TI symbols. For each group, a character string in which a serial number indicating each class belonging to the group is added to the symbol indicating the group is identification information.
[0057]
For the unlearned pattern, the serial number is common to all groups, and 0 is added after the symbol of the class.
[0058]
FIG. 7 shows a class (initial class) that can be recognized by the pattern recognition unit 2 in the initial state at the start of operation, identification information, and teaching patterns used for constructing dictionary information. The initial class is a tactile pleasant stimulus that appeals to the survival instinct of a virtual organism or a desire for acceptance, and the virtual organism in the present invention has evolved to recognize such signals and stimuli innately. Based on the assumption that
[0059]
There are two initial classes in the facial expression group: smile (FE1) and angry face (FE2). There are two initial classes in the morale group: gentle vocabulary (VE1) and angry morale (VE2). The class indicates that there are three types: stroke (TI1), strike (TI2), and presser (TI3).
[0060]
The state update unit 3 in FIG. 1 is means for holding the internal state information of the virtual organism and updating the internal state governing the emotional state of the virtual organism, for example, according to the identification information by the pattern recognition unit 2 described above. In step S13 of 2, the internal state update process is performed.
[0061]
In the virtual biological system according to this embodiment of the present invention, three types of parameters such as happiness level H, excitement level A, and preference level L are used as internal states that govern emotional states, as shown in FIG. The happiness parameter H is in the range of -1.0 to 1.0. Depending on the value in this range, the state of fear, fear, anxiety, relief, joy, jealousy, that is, anxiety and relief of virtual creatures. Represents the degree. If there is no special large stimulus, it will converge naturally.
[0062]
The degree of excitement parameter A is in the range of 0.0 to 1.0, and this value represents the degree of calm and excitement. If there is no extra large stimulus, it naturally converges to 0.0 (calm).
[0063]
Moreover, the parameter L of likability is in the range of −1.0 to 1.0, and the value within this range represents the degree of likes from dislike.
[0064]
The internal state parameters shown in FIG. 8 change as shown in FIG. 9, for example, according to each class when the initial class described above is detected by the pattern recognition unit 2. At this time, each class is classified as either a reward-type pleasant stimulus or a punitive-type unpleasant stimulus as shown in the figure. Are assigned angry faces, angry words, beatings and holdings. For example, the smile class (FE1), which is a pleasant stimulus corresponding to the affinity expression, gradually increases the values of the parameters of the happiness level H, the likability L, and the excitement level A in proportion to the duration of the stimulus. In addition, the anger speech (VE2), which is an unpleasant stimulus corresponding to a strong resentment, rapidly decreases the value of happiness level H and likability L in proportion to the number of stimuli, and rapidly increases the value of excitement level A. increase.
[0065]
Thus, the values of the parameters of the happiness level H, the excitement level A, and the likability L change as a result of receiving pleasant or unpleasant stimuli. Information defining how the internal state should be changed for each class is referred to as update information, and the state update unit 3 holds all the recognizable classes for each class.
[0066]
The output unit 4 is means for outputting a response by emotion expression and attitude expression based on the internal state and pattern input result updated by the state update unit 3 as described above, and is processed in step S14.
[0067]
The way of response in the output unit 4 is the property that the virtual creature has acquired in nature, and is set so that the response can be determined according to the internal state and the pattern input result as shown in FIG. For example, when the level of happiness reaches the level of jealousy, the virtual creature grows up with enchanting voices, expresses emotions such as joyful voices in the area of joy, and expresses people with good emotions. When it detects it, it tries to adjust its line of sight, and when it detects a person with bad emotion, it tries to turn its face.
[0068]
With the configuration described above, the virtual system according to this embodiment of the present invention recognizes the initial class, and, for example, responds that it is terrified when it is shouted (angry language) or scared (boiled). Functions as a virtual biological system.
[0069]
Next, a mechanism and features of the system of this embodiment for spontaneously learning a pattern will be described.
[0070]
The learning command generation unit 5 in FIG. 1 specifically, when the internal state appropriately updated by the state update unit 3 satisfies a predetermined condition, the happiness level H, the excitement level A, and the favorable level are caused by a pleasant stimulus or an unpleasant stimulus. When the absolute value of the sensitivity L exceeds a predetermined threshold value set when the value is a positive value or a negative value, a new input pattern that constitutes a situation that causes such an emotional state to be generated Or means for generating and outputting learning commands for all observed input patterns in order to learn again. Therefore, in step S15 of FIG. 2, it is determined whether or not to learn, and when any of the happiness level H, excitement level A, and likability L exceeds a predetermined threshold, it is determined that learning should be performed, The process moves to the pattern learning process of S16.
[0071]
Here, when comparing the present invention with the conventional example described above, in “Pichamin”, as described above, “learns a plurality of input patterns observed from the situation that caused the predetermined emotional state” It lacks functionality. Therefore, “Pichamin” selectively learns input patterns that indicate the occurrence of a situation that has a value (biological value) that is detected from the standpoint of the virtual creature in which the system should behave, for example, an emotional blow. I can't.
[0072]
For example, when the user A uses the virtual biological system according to the embodiment of the present invention, it is assumed that Mr. A gently strokes the pressure sensor while saying “good” while smiling.
[0073]
At this time, the patterns observed from the situation are the face image pattern of Mr. A, the voice spectrum pattern of “Yoshiyoshi”, and the pressure-sensitive pattern representing the stroke. If it is assumed that the face of Mr. A and the vocabulary “Yoshiyoshi” are not learned, the virtual biological system can recognize only TI1 with the smile FE1 and the gentle vocabulary VE1 and 撫 shown in FIG. That is, “FI0, FE1, VI0, VE1, TI1” is obtained as identification information by the pattern recognition unit 2.
[0074]
Based on “FE1, VE1, TI1” excluding the unlearned patterns FI0 and VI0, the pleasant state shown in FIG. Or the preference L increases, and the output unit 5 outputs an expression expressing a sense of joy, joy or relief. And when such a stimulus works cumulatively and the happiness level H and the likability L exceed a predetermined threshold, the learning command generation unit 5 generates a learning command.
[0075]
The pattern learning unit 6 is means for learning an input pattern that is observed in accordance with the learning command from the learning command generating unit 5, and learns the input pattern in step S16 when it should be learned in step S15. In step S15, if learning has not been done, the process returns to step S11 again to wait for input.
[0076]
The input pattern learning in step S16 is performed by one of the following two methods.
[0077]
If the identification information by the pattern recognition unit 2 represents an unlearned class (FI0, VI0 in the above example), it is assumed that the dictionary information that can recognize the input pattern is not stored in the pattern recognition unit 2. Dictionary information for recognizing this from the pattern is newly generated using principal component analysis. At the same time, the identification information is newly issued and stored in the pattern recognition unit 2.
[0078]
The newly issued identification information is given a number following the serial number of the existing class in the corresponding group. For example, FI1 is issued for a new class for identifying the person of Mr. A, and VI1 is issued for a new class for identifying the vocabulary “good”. When this new class is issued, the change amount of the internal state parameter value brought about by the known class “FE1, VE1, TI1” at that time is the amount by which each parameter is increased or decreased by the detection of the class, that is, the new class. Is stored in the state update unit 3 as update information for
[0079]
Therefore, the pattern learning unit 6 always monitors the difference information of the internal state parameter values before and after the generation of the learning command.
[0080]
On the other hand, when the identification information by the pattern recognition unit 2 represents a known class, the dictionary information of the class stored in the pattern recognition unit 2 is used so that the input pattern can be recognized better. It is strengthened using the learning subspace method disclosed in the above-mentioned document [1] and stored in the repeat pattern recognition unit 2.
[0081]
As a result, when a similar pattern is input from the next time, the pattern recognition unit 2 can recognize such a pattern more reliably and can respond accurately.
[0082]
As an example when Mr. A performs an action, when the absolute value of the emotion parameter value exceeds a predetermined threshold value, a learning command is generated by the learning command generation unit 5, and Mr. A's The face and the vocabulary “Yoshiyoshi” are learned. As a result, the system will be able to recognize the biologically valuable face of Mr. A, a person who seems to be happy, and the biologically valuable "goodness," a word that makes people happy. . On the other hand, no matter how much Mr. A's face and words are unlearned, the face and vocabulary of a person who does not give the biological value of pleasing or scaring will not be learned.
[0083]
By the learning command generation unit 5 and the pattern learning unit 6 described above, the virtual biological system according to this embodiment of the present invention can spontaneously learn an unlearned pattern and produce a virtual creature that develops pattern recognition ability. . In particular, the conditions that trigger learning are the detection of biologically valuable situations, that is, the generation of strong emotions such as “joyful” and “scary”, far more than learning patterns randomly or randomly. Biologically relevant learning ability can be realized in the present invention.
[0084]
As a result, the virtual biological system of this embodiment of the present invention not only provides a natural response to being scared when screaming, or engrossed when stroking, but it also improves by gaining experience afterwards. It functions as a virtual biological system that can return the action of being happy when it detects the acquaintance's face, or depressing when it detects an utterance, depending on the input pattern.
[0085]
The gist of the present invention is that the system can independently determine whether or not to perform pattern learning based on the emotional state driven by the recognition result of other already learned patterns input almost at the same time. is there. The function that pattern learning proceeds in accordance with the recognition result of other learned patterns input from the outside as described above is not present in “Pichamin” described above as a conventional example. This is because “Pechamin” recognizes only one group of patterns called speech vocabulary.
[0086]
The functions described above can be realized not only by hardware in which components having respective functions are combined but also as software.
[0087]
The present invention also provides a computer-readable recording medium storing a program for causing a computer to execute a predetermined procedure, a program for causing a computer to function as a predetermined means, or a program for causing a computer to realize a predetermined function. It can also be implemented as a recording medium.
[0088]
As shown in FIG. 11, information representing the pattern recognition method and pattern registration method according to the present invention, for example, a program is recorded on a recording medium 31, and the recorded information is stored in the computer device 32 via the recording medium 31. It is also possible to apply to the computer apparatus 34 via the communication line 33.
The present invention is not limited to the embodiments described above, and can be implemented with various modifications within the scope of the technical idea.
[0089]
【The invention's effect】
As described above, according to the present invention, a virtual organism system and a virtual organism capable of performing spontaneous pattern learning that eliminates inappropriate patterns by having a more appropriate means for determining when to learn a pattern. A pattern learning method in the system can be obtained.
[Brief description of the drawings]
FIG. 1 is a diagram showing a functional block configuration example of a virtual biological system according to an embodiment of the present invention.
FIG. 2 is a diagram showing a processing flow in the virtual biological system according to the embodiment of the present invention.
FIG. 3 is a view showing an arrangement example of various input means in the robot-type virtual biological system according to the embodiment of the present invention.
FIG. 4 is a view showing an arrangement example of various input means in the virtual biological system in the screen display format according to the embodiment of the present invention.
FIG. 5 is a view for explaining groups that can be recognized and their constituent dictionaries in the virtual biological system according to the embodiment of the present invention;
FIG. 6 is a diagram for explaining groups that can be recognized in the virtual biological system according to the embodiment of the present invention and formats of their identification information.
FIG. 7 is a diagram for explaining classes and their teaching patterns that can be recognized in the initial state in the virtual biological system according to the embodiment of the present invention.
FIG. 8 is a diagram for explaining an internal state parameter governing an emotion state in the virtual biological system according to the embodiment of the present invention.
FIG. 9 is a diagram for explaining changes in internal state parameters by stimulation in the virtual biological system according to the embodiment of the present invention.
FIG. 10 is a diagram for explaining a response according to an internal state parameter in the virtual biological system according to the embodiment of the present invention.
FIG. 11 is a view for explaining another embodiment in the case where the virtual biological system of the present invention is realized by being transferred by a recording medium.
[Explanation of symbols]
DESCRIPTION OF SYMBOLS 1 ... Input part, 2 ... Pattern recognition part, 3 ... State update part, 4 ... Output part, 5 ... Learning command production | generation part, 6 ... Pattern learning part, 11 ... Robot, 12 ... TV camera, 13 ... Microphone, 14 ... Pressure sensor, 21 ... Computer, 22 ... Display virtual creature, 23 ... TV camera, 24 ... Microphone 25 ... mouse, 26 ... left button, 31 ... recording medium, 32, 34 ... computer device.

Claims (12)

An input unit for inputting pattern information from a plurality of types of signal sources, a pattern recognition unit for recognizing the input pattern information based on dictionary information, and a state for updating an emotional state based on a recognition result by the pattern recognition unit In a virtual biological system comprising update means and output means for outputting a response according to the emotional state,
When the emotional state satisfies a predetermined condition, further comprising learning means for learning and new dictionary information for recognition of the pattern information, and the amount of updating the emotion state by detection of said pattern information Virtual biological system characterized by that. An input process for inputting pattern information from a plurality of types of signal sources, a pattern recognition process for recognizing the input pattern information based on dictionary information, and a state for updating an emotional state based on a recognition result by the pattern recognition means A pattern learning method in a virtual biological system, comprising: an update process; and an output process for outputting a response according to the emotion state ,
It further has a pattern learning process for learning new dictionary information for recognizing the pattern information and an amount of updating the emotional state by detecting the pattern information when the emotion state satisfies a predetermined condition. A pattern learning method in a virtual biological system characterized by the above. The virtual biological system according to claim 1, wherein the predetermined condition is that the intensity of the emotional state exceeds a predetermined threshold. 3. The pattern learning method in the virtual biological system according to claim 2, wherein the predetermined condition is that the intensity of the emotional state exceeds a predetermined threshold. Input means for inputting pattern information from one or more signal sources;
Pattern recognition means for recognizing the pattern information input by the input means for each group based on dictionary information classified into a plurality of groups;
Pattern learning means for learning new dictionary information for recognizing the pattern information, and an amount of updating the emotional state by detecting the pattern information ;
State update means for updating the emotion state based on the recognition result by the pattern recognition means;
A learning command means for giving a learning command to the pattern learning means when the emotional state satisfies a predetermined condition;
Virtual biological systems, characterized by comprising output means for outputting a response in accordance with the emotional state. Input means for inputting pattern information from one or more signal sources;
Pattern recognition means for recognizing the pattern information input by the input means for each group based on dictionary information classified into a plurality of groups;
Pattern learning means for learning new dictionary information for recognizing the pattern information, and an amount of updating the emotional state by detecting the pattern information ;
State update means for updating the emotional state by changing these numerical values of the emotional state represented by a plurality of numerical values based on the recognition result by the pattern recognition means;
A learning instruction means for issuing a learning instruction to the pattern learning means when a plurality of numerical values representing the emotional state have reached a predetermined size;
A virtual biological system comprising output means for outputting a response according to the emotional state. Input means for inputting pattern information from one or more signal sources;
Pattern recognition means for recognizing a plurality of pattern information input within a predetermined time by the input means for each group based on dictionary information classified into a plurality of groups;
Pattern learning means for learning new dictionary information for recognizing the pattern information, and an amount of updating the emotional state by detecting the pattern information ;
State update means for updating the emotional state by changing these numerical values of the emotional state represented by a plurality of numerical values based on the recognition result by the pattern recognition means;
A learning instruction means for issuing a learning instruction to the pattern learning means when a plurality of numerical values representing the emotional state have reached a predetermined size;
Virtual biological systems, characterized by comprising output means for outputting a response in accordance with the emotional state. The virtual biological system according to claim 6 or 7, wherein the plurality of numerical values representing the emotional state are numerical values representing at least happiness, excitement, and likability. 8. The virtual information according to claim 6, wherein the pattern information is recognized for at least one of a group of people and vocabulary and at least one of a group of facial expressions, vocabulary, and force sense. Biological system. The pattern information is divided into groups of persons, facial expressions, vocabulary, vocabulary and force sensation , and the emotional state is expressed by the pleasantness and discomfort of the stimulus recognition results for the facial expression, vocabulary and force sense groups. 8. The virtual biological system according to claim 6 or 7, wherein numerical values of happiness, excitement, and likability that express the value are changed. Inputting pattern information from one or more signal sources; and
A pattern recognition step for recognizing a plurality of pieces of pattern information input within a predetermined time by this step for each group based on dictionary information classified into a plurality of groups;
Based on the recognition result of this pattern recognition step, a state update step of updating the emotional state by changing these numerical values of the emotional state represented by a plurality of numerical values,
When at least one of a plurality of numerical values representing the emotional state reaches a predetermined size , new dictionary information for recognizing the pattern information, an amount for updating the emotional state by detecting the pattern information, A learning command step for commanding to learn
A pattern learning step for learning the new dictionary information and the update amount in accordance with a command performed by the learning command step;
Pattern learning method in a virtual biological systems, characterized by an output step of outputting a response in accordance with the emotional state. The pattern information is divided into groups of persons, facial expressions, vocabulary, vocabulary and force sensation , and the emotional state is expressed by the pleasantness and discomfort of the stimulus recognition results for the facial expression, vocabulary and force sense groups. The pattern learning method in the virtual biological system according to claim 11, wherein numerical values of happiness level, excitement level, and likability that express the value are changed.

Images