JP5624100B2 - Artificial emotion learning apparatus and method - Google Patents

Description

  The present invention relates to an artificial emotion learning apparatus and method, and more particularly, to an artificial emotion learning apparatus and method for a robot capable of generating an emotion by a method desired by a user through learning.

  In general, the emotion of a robot is limited to an emotion generated at a specific position in an emotional space having a predetermined number of emotions by input from a sensor.

  FIG. 1 is a schematic diagram showing a general robot emotion expression method.

  In order to express emotions as a robot, the emotion value of the current robot must be calculated. Emotions are rarely defined by a single emotion, such as joy or sadness. Even if humans feel joy now, some other emotions such as surprise emotion and anger emotion are reflected. That is, the emotional expression is a result of reflecting complex detailed emotions. Therefore, when embodying realistic emotional expression through the robot, the emotional value applied to the robot also has various detailed emotions such as happiness, sadness, surprise, anger, etc. It can be reflected and expressed as a vector.

  As shown in FIG. 1, generally, a two-dimensional or three-dimensional fixed dimensional space is used to express a robot's emotion, and the emotion and its position are placed at a fixed position on the fixed dimensional space. Map emotional expressions that correspond to emotions. The emotion value can be expressed and calculated as a vector value corresponding to a certain position in space.

  That is, if emotions are mapped to various points on the vector space, and emotion expressions corresponding to each emotion are mapped one-to-one, then given emotion vectors, A method has been adopted in which one of the emotions mapped in the vector space is selected from the emotions closest in distance, and the emotion mapped one-to-one with the emotion is expressed.

  In other words, since there is a limit in manually mapping emotions and emotion expressions corresponding to the emotions to countless coordinates in the vector space, the conventional method of FIG. After mapping the emotions that correspond to the emotions and the emotional expression behaviors that correspond to the emotions, the emotions of the coordinates closest to the emotion vector determined by reflecting the emotional state of the robot are selected, and the emotions are expressed accordingly It is.

  For example, it is set to express an emotion value 1 {joy 1, sadness 0, surprise 0, anger 0} at coordinates 1 on a four-dimensional vector space, and emotion value 2 {joy 3/4, sadness 1/4, If surprise 0, anger 0} and emotion value 3 {joy 3/4, sadness 0, surprise 1/4, anger 0} are closer to coordinate 1 than coordinates expressing other emotions, emotion values 1, 2, 3 All perform the emotional expression set at the coordinate 1.

  As described above, in the conventional method, although the emotion value actually generated internally is different, the emotion value to be selected is only one of the most similar emotion values mapped to the coordinate 1. Since the emotion expression operation is selected based on the emotion value of the same coordinate selected, the form generally expressed through the expression organization is the same.

  For more detailed explanation, FIG. 2 is used. FIG. 2 is a schematic diagram for explaining a process of generating a robot emotion from a conventional emotion state input value. In FIG. 2, a one-dimensional emotion coordinate system is used for convenience of explanation.

  In the emotion coordinate system, it is assumed that a basic emotion of joy is arranged at coordinate 1, and a basic emotion of sadness is arranged at coordinate-1. If the input value is 0.3, the basic emotion closest to 0.3 is extracted. In the case of FIG. 2, since 0.3 is closer to 1 than −1, the basic emotion of joy is extracted. Since the basic emotion of joy is coordinate 1 in the emotion coordinate system, the emotion of the robot eventually becomes coordinate 1.

  According to the emotion generation method as described above, even if the input value is 0.5, the emotion of the final robot is coordinate 1 as in the case where the input value is 0.3.

  Therefore, in the case of an emotion expression device that expresses eyes, mouth, gestures, etc. by inputting the robot emotion generated by the emotion generation method described above, the emotion state input value is different from 0.3, 0.5, etc. However, the same coordinate 1 is expressed as the emotion of the robot. Therefore, most conventional robots express the same emotion even if the input values are different.

  However, since the conventional robot emotion generation device always generates the same emotion in response to the same input in this way, if an undesired emotion appears depending on the user's tendency, it becomes impossible to sympathize with the robot, and the interest in the robot is halved. There was a problem of doing.

Miwa H., et al., "Effective emotional expressions with expression humanoid robot WE-4RII: integration of humanoid robot hand RCH-1," In Proceedings of 2004 IEEE / RSJ International Conference on Intelligent Robots and Systems (IROS 2004), vol .3, pp. 2203-2208, 2004. C. Breazeal, "Function meets style: insights from emotion theory applied to HRI," IEEE Transactions on Systems, Man and Cybernetics, Part C: Applications and Reviews, Vol.34, no.2, pp.187-194, 2004. Park Jeong-Woo, Lee Hui-Sung, Jo Su-Hun, and Chung Myung-Jin, "Dynamic Emotion Model in 3D Affect Space for a Mascot-Type Facial Robot," In Proceedings of the 2nd Korea Robot Society Annual Conference, 2007 . Christian Becker, Stefan Kopp, and Ipke Wachsmuth, "Simulating the Emotion Dynamics of a Multimodal Conversational Agent," In Proceedings Tutorial and Research Workshop on Affective Dialogue Systems, pp.154-165, 2004.

  In order to solve the problems as described above, an artificial emotion learning device capable of learning a robot capable of generating a user's desired emotion by changing the distribution of emotion arranged in the emotion space through learning by the user, and Its purpose is to provide a method.

  In order to solve the above problems, an artificial emotion learning apparatus according to the present invention includes an internal state calculation unit that detects external environment information and calculates an internal state vector value corresponding to one point of an internal state coordinate system, and the internal state coordinates. The emotion determination processing unit for generating an emotion value in each emotion value group indicated by the coordinates of the internal state input value of the robot on the system as the emotion of the robot, and the emotion value determined by the emotion determination processing unit, The internal state coordinate system is configured to receive an action generation unit that defines an action of the robot and expresses emotion, an emotion value determined from the emotion determination processing unit, and emotion type information input from a user input unit. An emotion learning unit that generates feedback information related to an emotion value corresponding to one point, and changes the emotion probability distribution of the internal state coordinate system using the feedback information; and the emotion learning unit And a user input unit for inputting information about the type of information. Here, it is preferable that the emotion probability distribution is a single Gaussian distribution having a weight value, a median value, and a variance, or a Gaussian distribution obtained by combining a plurality of the Gaussian distributions.

  In particular, the emotion learning unit may search for a Gaussian distribution corresponding to the feedback information using a matching function and change the Gaussian distribution.

  On the other hand, in order to solve the above problem, the artificial emotion learning method according to the present invention is such that the internal state calculation unit senses external environment information and calculates an internal state vector value corresponding to one point of the internal state coordinate system. An emotion determination processing unit including a state generation step and an emotion model in which an emotion probability distribution for j types of emotions is formed on an n-dimensional internal state coordinate system; and an input value of the internal state calculation unit A step of determining an emotion value of the robot by matching with the internal state coordinate system; a step of expressing an emotion by determining an operation of the robot by an emotion value determined by the emotion determination processing unit; The learning unit receives the emotion value determined from the emotion determination processing unit and the emotion type information input from the user input unit, and relates to the emotion value corresponding to one point of the internal state coordinate system. It generates feedback information that includes a step of changing the emotion probability distribution of the internal state coordinate system by using the feedback information.

  ADVANTAGE OF THE INVENTION According to this invention, the emotion which a user desires can be produced | generated by changing the distribution of the emotion arrange | positioned on emotion space through a user's learning. Therefore, it is possible to increase the interest in the robot by showing emotions that can be sympathized with the user's tendency.

It is the schematic which shows the emotion model of a general robot. 6 is a diagram illustrating an embodiment of a process of generating a robot emotion from an emotion state input value in an emotion space according to a conventional technique. It is a block diagram which shows the artificial emotion learning apparatus by this invention. It is a block diagram which shows each structure of the artificial emotion learning apparatus by this invention, and the space between each structure. It is the schematic which shows the emotion probability distribution employable for one Embodiment of this invention. 3 is a diagram illustrating an emotion model having an emotion probability distribution in which two Gaussian distributions of the same emotion are combined on a one-dimensional emotion space according to an embodiment of the present invention. 3 is a diagram illustrating an emotion model having two emotion probability distributions for other emotions on a one-dimensional emotion space according to an embodiment of the present invention. It is a conceptual diagram which shows the internal state calculation apparatus by this invention. It is a conceptual diagram which shows the internal state calculation method by this invention. It is the schematic which shows the emotion model by one Embodiment of this invention. It is the schematic which shows the emotion model changed by learning by one Embodiment of this invention. 3 is a flowchart of an artificial emotion learning method according to an embodiment of the present invention.

  Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings.

  3A and 3B are block diagrams showing an artificial emotion learning apparatus 1000 using a generative probability model to which the present invention is applied.

  First, the overall configuration of the artificial emotion learning apparatus 1000 will be described with reference to FIGS. 3A and 3B.

  As shown in FIG. 3A and FIG. 3B, the artificial emotion learning apparatus 1000 learns artificial emotions based on a generative probability model, so that four spaces are an external input space 910, an internal state space 920, an emotion space 930, and an operation. The space 940 is configured by an internal state calculation unit 100, an emotion determination processing unit 200, an action generation unit 300, an emotion learning unit 400, and a user input unit 500, which are four components that perform a mapping process and the like.

  First, an external input space (External Input Space) 910 includes a vector space using information sensed externally such as facial expressions, gestures, and tone.

  The internal state space 920 is composed of the five-factor model of personality, and the five factors of personality are openness to experience, conscientiousness, extroversion ( It is formed with extraversion, agreeability, and neuroticism.

  The internal state space 920 can also be configured as a one-dimensional or higher emotion space 930 using one or more of the five personality factors, and the emotional state is obtained by combining one or more of the five personality factors with the other factors. The space 930 can also be configured. Further, the emotion space 930 may be configured by any factor, such as an emotion space 930 configured using other factors without using one of the five factors of personality, and any of them can be applied.

  The internal state space 920 is a concept introduced to construct a basic internal state for a robot having a personality and emotion to react to a specific external input. More specifically, even if there is the same external input or stimulus, each human being reacts according to the background and environment that has grown, that is, the emotions generated are different, and the ultimate is based on such different emotions. Express your own emotions to the outside.

  For example, when the same external stimulus is input to a person who has been exposed to a violent environment for a long time and a person who is not, the former is likely to have an internal state, that is, a very sharp personality, While the likelihood of generating an emotion based on such a state is high, the latter is likely not.

  The emotion space 930 is composed of a vector space based on the magnitude of emotion such as happiness, surprise, sadness, love, dislike, fear, and anger.

  The behavior space 940 is composed of a vector space based on the amount of unit behavior indicating the smallest motion such as a mouth, mouth opening, etc. of each emotion expression organization such as the mouth, eyes, and head. .

  3A and 3B calculates an internal state as an internal process, maps it from the external input space 910 to the internal state space 920, and defines the internal state space 920 and dimensions (Dimension). To do. The internal state means a point in the internal state space 920 that uses an external input.

  Prior art internal state space 920 was arbitrarily modified despite being based on psychological studies. Also, since the result of the internal process is a point on the internal state space 920, the internal process must be connected to the state space. However, the conventional technique has a problem in that an internal process is arbitrarily connected to the internal state space 920 according to a user's experience.

  In the present invention, the internal state space 920 is constructed based on the five-factor model of personality, which is a study of psychology. The five-factor model of personality is introduced to explain the human personality psychologically, and is well known to those skilled in the art, so a detailed description is omitted in the present invention.

  The internal process is configured using NEO-PI-R (the Revised NEO Personality Inventory), which is a method of measuring a personality five-factor model. Of course, as described above, the internal process may be configured using a different one from NEO-PI-R. NEO-PI-R will be described later.

  The emotion determination processing unit 200 in FIGS. 3A and 3B selects an emotion vector e (k) mapped to the emotion space 930 in the internal state space 920 as an emotion determination process.

  In the prior art, since the internal state space 920 is arranged in a plurality of fixed areas for classifying emotions, it is difficult to reflect differences in age and cultural background. In addition, there was a problem that the definition and number of basic emotions differed for each psychological study.

  However, the present invention sets an emotion probability method using a mixed Gaussian distribution (GMM, Gaussian Mixture Model) on the emotion space 930, and updates the emotion probability by user feedback. In the present invention, the emotion vector e (k) is selected through probability values distributed over all emotions.

  The motion generation unit 300 in FIGS. 3A and 3B maps the emotion space 930 to the motion space 940 in order to generate a motion of the robot as a behavior generation process.

  In the past, robot behavior and actions were selected using a predetermined 1: 1 mapping between actions and emotions. Therefore, the existing robot always expressed the same movement when the same external input or the like is given. Moreover, since one generated motion reflects only one emotion, the robot has to express only simple and fragmented motion.

  In the present invention, in order to determine the most probable motion of the robot, the concept of a motion generation process for generating various motions in consideration of the unit motion and the probability emotion value is introduced.

  Hereinafter, the generation of the internal state in the emotion determination process based on the generative probability model will be described.

  First, as shown in FIG. 3A and FIG. 3B regarding the internal state generation, the artificial emotion learning apparatus 1000 using the generative probability model of the present invention has two processes: an internal process and an emotion determination process having a learning scheme. Includes one process.

  The internal process maps the external input space 910 to the internal state space 920, and generates an internal state at one point existing in the internal state space 920. The internal process is modeled using NEO-PI-R (Revised NEO Personality Inventory), a method developed to measure a five-factor model of personality.

  As described above, the five-factor model of personality is a model for explaining an individual's personality, and includes the following five aspects of personality for explaining the personality of an individual. The five factors are openness to experience, conscientiousness, extraversion, agreeableness, and neuroticism.

  NEO-PI-R is a psychological personality test method and is used as a measurement method for a five-factor model of personality. The NEO-PI-R experiment includes six subelements (faces) known to be the facets of each factor in the personality five-factor model, and these can be measured.

In NEO-PI-R, it is formed of 6 subelements per upper element as follows.
・ Openness to experience: fantasy, aesthetics, feelings, actions, ideas, values
・ Conscientiousness: competence, order, dutifulness, achievement striving, self-discipline, deliberation
・ Extraversion: warmth, sociality, assertiveness, activity, excitement seeking, positive emotion
・ Agreeableness: trust, straightforwardness, altruism, compliance, modesty, tender mindedness
・ Neuroticism: anxiety, hostility, depression, self-consciousness, impulsiveness, stress vulnerability (vulnerability to stress)

  In FIG. 3A and FIG. 3B, u (k) is defined as an external input vector, and elements of the external input vector u (k) indicate external input states such as facial expressions, gestures, and tone. . The external input vector u (k) exists in an S-dimensional external input space 910 having S external inputs.

  x (k) is defined as an internal state vector, and the elements of the internal state vector x (k) are based on a five-factor model of character, and are open, honest, It represents the five factors of extroversion, affinity and neurotic tendency.

  The external input vector u (k) becomes an input of the internal process, and the internal state vector x (k) becomes an output of the internal process. If there is no external input for a predetermined time, the internal state falls within 0 where the emotional state is neutral.

  The emotion decision process is a process of mapping from the internal state space 920 to the emotion space 930, and generates an emotion vector e (k) that is one point existing on the emotion space 930. In the present invention, an emotion probability distribution distributed on the internal state space 920 is set, and the emotion probability distribution is updated by learning.

  e (k) is defined as an emotion vector, and an element constituting the emotion vector e (k) represents the probability size of the emotion indicated by the corresponding element. The emotion space 930 has a J dimension, where J represents the number of emotions.

  The internal state vector x (k) is input to the emotion determination process, and the emotion vector e (k) is output from the emotion determination process. When the internal state vector x (k) is input to the emotion determination process, the emotion probability is determined using the emotion probability distribution distributed in the internal state space 920, and the emotion vector e (k) On the other hand, it is determined to have a probability.

  The motion generation unit 300 performs emotional expression using the emotion vector e (k) determined by the emotion determination processing unit 200. The motion generation unit 300 can perform robot facial expression processing, voice output processing, motion processing using a robot arm, legs, torso, and the like. On the other hand, since the motion generation unit 300 is not the main aspect of the present invention, detailed description thereof is omitted.

The emotion learning unit 400 receives the emotion vector e (k) determined from the emotion determination processing unit 200 and the emotion type information e _d (k) input from the user input unit 500, and the emotion space 930 Feedback information corresponding to the emotion vector e (k), which is one point existing above, is generated, and the emotion probability distribution is changed using the feedback information. That is, feedback information regarding an emotion value corresponding to one point of the internal state coordinate system is generated, and the emotion probability distribution of the internal state coordinate system is changed using the feedback information.

  The user input unit 500 inputs emotion type information to the emotion learning unit 400.

  FIG. 4A is a schematic diagram showing an emotion probability distribution that can be employed in an embodiment of the present invention.

  First, referring to FIG. 4A, FIG. 4A is a coordinate indicating the internal state space 920, based on a psychological FFM (Five Factor Model), a five-dimensional internal state (openness to experience, sincerity). Conscientiousness, extraversion, agreeableness, neuroticism space, and the conical shape of the same texture shown in FIG. Is shown. That is, one conical shape indicates one emotion probability distribution. As shown in FIG. 4A, a probability distribution curve representing an emotion of G1 is drawn at a position close to the basic internal state axis of the robot of outwardness and affinity.

  Here, one emotion distribution is expressed as one generative probability model having a plurality of Gaussian distributions G1 or Gaussian modes, and the emotion space 930 is expressed as a plurality of generative probability models forming a volume space. The At this time, the volume space is represented by a plurality of emotion values that are the largest in the coordinates of the basic emotion and gradually decrease as the distance from the coordinates of the basic emotion increases. The emotion value is a size representing emotion. For example, assuming that there is purely a feeling of joy, the magnitude of joy differs even with the same joy. The emotional value is used to express the magnitude of joy.

  The basic emotion is an emotion given to the robot by the user and can be defined as joy, sadness, anger, surprise, fear, and the like. These basic emotions are set on the internal state space coordinate system of the robot.

  Various emotion value groups are set, and the shape of the volume space generated by these emotion value groups appears as the personality and propensity of the robot. For example, as shown in FIG. 4A, cones with the same texture represent the same emotion, and the form of each cone is different for each robot, that is, this becomes the uniqueness of the robot. In addition, it is advantageous for natural emotional expression to increase the emotion value at the basic emotion coordinates at the time of setting the emotion value group, and gradually reduce the emotion value as it moves away from the basic emotion coordinates. There is no need to be done.

  On the other hand, the emotion value group is generated based on the internal state input value. In order for the robot to express emotions, external stimuli are necessary at an early stage. Such a stimulus is detected by an emotion-related sensor provided in the robot, and the detection result of the sensor is processed into data that can be expressed on the coordinate system of the internal state space 920 that the robot has. Data processed in this way can be an internal state input value. Also, dynamic emotion expression can be achieved by changing the emotion value group according to the internal state input value.

  The internal state calculation unit 100 senses external environment information, that is, an external stimulus by an emotion-related sensor provided in the robot, and processes it into data that can be expressed on a coordinate system. Here, external stimuli are voice recognition (meaning of words, loudness, etc.), video recognition (human recognition, object recognition, hue recognition, etc.), various sensor values (touch sensor position value, intensity, All stimuli that humans can recognize, such as distance traveled, temperature, humidity, etc.) can be used.

  In the case of a robot to which learning ability is given, emotion coordinates of pleasure are changed, added, or deleted by learning. For example, as shown in FIG. 4B, in addition to the pleasure coordinate 1, a pleasure coordinate 0.5 is added by learning. In such a case, since there are two same emotions, they are combined and shown in one emotion value group. This will be described in detail later in the emotion learning unit 400.

  FIG. 4B shows a case in which each emotion value group forms a Gaussian curve centered on the coordinates of basic emotions, and a Gaussian curve with pleasure coordinates 1 and a Gaussian curve with pleasure 0.5 newly added by learning. Shows a state of being merged by GMM (Gaussian Mixture Model).

  At this time, the weight of the Gaussian distribution must always satisfy 1 for each emotion. For example, if only one Gaussian distribution exists for the A emotion, the weighted value of the Gaussian distribution is 1. Similarly, if there are two Gaussian distributions for the A emotion, the weights of the two Gaussian distributions are different, but the sum must always satisfy 1.

  In this way, on the coordinate system of the internal state space 920, a set of emotion values of each emotion value group indicated by the coordinates of the internal state input value of the robot can be generated as the emotion of the robot.

  On the other hand, in the present invention, the emotion value group is indicated by a Gaussian curve, but the emotion value group may be indicated by various curves that satisfy the above-described volume space conditions, such as a triangle. That is, as yet another embodiment, the emotion probability distribution is formed into a distribution having various shapes (for example, a circular distribution, a square distribution, a triangular distribution, a semicircular distribution, etc.) that is not a Gaussian distribution.

  As described above, distributions other than the Gaussian distribution can be applied. In this case, a mathematical expression different from the Gaussian distribution mathematical expression can be applied, but the principle is the same. In this specification, a detailed description of mathematical expressions relating to other distributions is omitted.

  However, by using the same Gaussian curve of dispersion around the median (emotion value for basic emotions, etc.), by creating a uniform distribution of emotion values around the reference emotion in the robot's emotion coordinate system, Although it is possible to easily calculate the emotion vector e (k) indicating the emotion of the robot, the present invention is not necessarily limited to this.

  The emotion determination processing unit 200 applies an emotion state input value as an input value to the emotion model, and calculates an emotion value as an output value.

  FIG. 4C is a diagram illustrating an emotion model having two emotion probability distributions with respect to other emotions on a one-dimensional emotion space according to an embodiment of the present invention. It is the schematic for demonstrating. For convenience of explanation, as shown in FIG. 4C, the emotion model is described using an emotion curve located on the one-dimensional internal state space 920.

  In FIG. 4C, a sadness coordinate −1 which is a basic emotion different from the joy coordinate 1 which is a basic emotion is given on a one-dimensional coordinate, and an emotion probability distribution which forms a Gaussian curve for each basic emotion. Has been granted.

  If the internal state input value is 0.3, the coordinate 1 of joy is the emotion of the robot, but according to the emotion generation unit 130 of the present embodiment, each coordinate represented by 0.3 {4, 1}, which is a set of emotion values in the emotion value group, is the robot's emotion. Of course, for such an expression, the element order of the set must be set in advance. In the present embodiment, a case is shown where {emotion value of joy, emotion value of sadness} is set.

  According to the present embodiment, instead of simply expressing the emotion of the robot with the coordinate 1 representing joy, the emotion of joy is 4 and the emotion of sadness is 1 It can be generated with high efficiency. At this time, since the compound emotion changes depending on the shape of the Gaussian curve, the Gaussian curve forms the character of the robot.

  Also, depending on the dimension, when one internal state input value represents multiple input values in one emotion probability distribution, the largest value, the smallest value, or a value close to the average is selected to extract one emotion value. May be. As a result, when the internal state coordinate system is expanded in multiple dimensions, only one input value represented by the internal state input value may be extracted per internal state. On the other hand, in consideration of the convenience of the robot emotion expression device that receives the robot emotion, the emotion value of the set, which is each element of the complex emotion, can be expressed stochastically. Probabilistic expression means expressing emotion values as a percentage or a fraction with a total of one. For example, in the above example, if the robot's emotion is {4, 1}, it will be {80, 20} when expressed as a percentage, and {4/5, 1/5} when expressed as a fraction where the total is 1. Become.

Next, the emotion learning unit 400 is a configuration for learning an emotion model. As an input value, an emotion vector e (k) that is an output value of the emotion determination processing unit 200 and emotion type information e through the user input unit 500 _d (k) is input, and feedback information for changing the weighted value, median value, and variance of each Gaussian distribution on the emotion model is output as an output value. That is, learning means changing the Gaussian distribution distributed on the emotion space 930, and means changing the weighted value, median value, and variance value of the Gaussian distribution. In particular, the emotion learning unit 400 searches for a Gaussian distribution corresponding to feedback information using a matching function and changes the Gaussian distribution.

  As a preferred embodiment, when an input saying “I love you with open arms” comes into the artificial emotion learning apparatus 1000 according to the present invention at k hours, the user who sees this says “happy” Education is possible. If “happiness” is input to the emotion learning unit 400 through the user input unit 500, the Gaussian distribution around the state vector is learned by the set method.

  Hereinafter, the operation of the artificial emotion learning apparatus 1000 of the present invention will be described using mathematical expressions.

(1) Definition of Internal State Vector x (k) FIGS. 5A and 5B illustrate an internal process using NEO-PI-R, and the process of FIGS. 5A and 5B is performed in the internal state space 920. Modeled using NEO-PI-R, a method for measuring a five-factor model of the character used.

  Referring to FIGS. 5A and 5B, the internal state calculation is performed in the order of face value vector calculation for personality factors, personality factor value calculation, and internal state calculation.

Referring to FIGS. 5A and 5B, the external input vector u (k) ε {0, 1} ^s is determined by the external input, where S represents the number of external inputs. The value of the i-th element in the external input vector u (k) represents whether the i-th sensor is operating, and represents 1 when it is in operation, and 0 otherwise. However, iε {1, 2,..., S}.

  The input vector is processed by a dynamic system with three personality parameters.

Facet parameter 151 (first character): Γ _o , Γ _c , Γ _e , Γ _a , Γ _n ∈ R ^{6 × s}
Personality factor parameter 152 (second personality): β _o , β _c , β _e , β _a , β _n εR
Durability parameter 153 (third character): α _o , α _c , α _e , α _a , α _n ∈ [0, 1]
The subscripts “o”, “c”, “e”, “a”, and “n” of each parameter represent openness, honesty, extroversion, affinity, and neurotic tendency.

  The Durability (persistence) parameter 153 (third personality) represents the sustainability of each personality factor according to the personality five-factor model, and the personality factor parameter 152 (second personality) represents the degree of sensitivity of each personality factor.

  The Facet parameter 151 (first personality) represents the degree of sensitivity of the six sub-elements included in each personality factor. The Facet parameter 151 (first personality) can be 30 in total because the five personality factors each include six subelements.

  For example, the facet parameter 151 (first personality) for openness, which is one of personality factors, can be expressed as follows.

_{Γ o = [γ o1, γ} o2, γ o3, γ o4, γ o5, γ o6] is a ^{_{T, γ o1, γ o2,}} γ o3, γ o4, γ o5, a γ _o6 ∈R ^s.

  The Face parameter 151 (first character) of the other four factors can be expressed in the same manner.

  Then, in connection with the calculation of the facet value vector for the personality factor, the input vector is used for the openness, integrity, extroversion, affinity and neurotic tendency using the corresponding Facet parameter 151. Is converted to a face value vector of five factors of sex.

For example, f _o (k) (where f _o (k) εR ⁶ ), which is a facet value vector for openness, is defined as follows.

f _o (k) = Γ _o u (k), and f _o (k) = [f _o1 (k), f _o2 (k), f _o3 (k), f _o4 (k), f _o5 (k) ), F _o6 (k)] ^T.

The face value vectors f _c (k), f _e (k), f _a (k), and f _n (k) of the other four personality factors are determined in the same manner.

  Then, in connection with the calculation of the personality factor value, the facet value vector can be calculated as a personality factor value by combining the six facet values.

For example, g _o (k) (where g _o (k) εR), which is a factor value for openness, is defined as follows.

であり、１_６は、１で満たされた６×１ベクトルである。

And ₁₆ is a 6 × 1 vector filled with 1.

The values for the other four personality factors, g _c (k), g _e (k), g _a (k), and g _n (k), are also determined in the same manner.

  Finally, in relation to the internal state calculation, it is converted into an internal state value using the Durability parameter 153, the personality factor parameter 152, and the past internal state value.

  For example, o (k + 1) for openness is defined by the following equation of state.

o (k + 1) = α _o o (k) + β _o g _o (k)

  c (k + 1), e (k + 1), a (k + 1), and n (k + 1) are determined by the same method.

That is, by adding the calculated is beta _o g _o the current personality factors value (k), a value obtained by multiplying the internal status of the immediately preceding and alpha _o is a Durability parameter 153 indicating the persistence of the internal state, the current Calculate the internal state value of.

  In relation to the overall internal state value, the overall internal state dynamics can be expressed by the following state equation.

x (k + 1) = Ax (k) + Bu (k)
x (k) = [o (k) c (k) e (k) a (k) n (k)] ^T ∈ R ⁵

  Moreover, it is not restricted to using all the five upper elements by NEO-PI-R, You may use 1 or more as needed. Further, it is not necessary to use all six lower elements included in the upper element, and one or more lower elements may be used as necessary. In this case, the dimension value of the internal state varies.

(2) Definition of emotion vector e (k), which is an emotion value of the robot The final emotion is not expressed as one emotion value, but is generated as a composite emotion that reflects the probability values of multiple emotion values . This complex emotion becomes an emotion vector e (k). If J emotions are used, it can be expressed as the following robot emotion equation 3.

ここで、Ｊは、基本感情の数であり、ｅｊは、ｊ番目の基本感情の感情値である。例えば、ｅ１＝ｈａｐｐｉｎｅｓｓ、ｅ２＝ｓｕｒｐｒｉｓｅ、ｅ３＝ｓａｄｎｅｓｓ、ｅ４＝ｌｏｖｅなどである。［ ］Ｔは、転置行列であって表現方式が一つであるので、削除できる。

Here, J is the number of basic emotions, and ej is the emotion value of the jth basic emotion. For example, e1 = happines, e2 = surprise, e3 = sadness, e4 = love, and the like. [] T is a transposed matrix and has only one expression, and can be deleted.

(3) Setting of emotion probability distribution P (x (k) | ej) The probability distribution for each basic emotion is arranged on the internal state coordinate system. The probability distribution for each basic emotion uses GMM (Gaussian Mixture Model). If this is used, the distribution of each basic emotion can be made dynamic, and each emotion value can be calculated independently. An emotion probability distribution (emotion distribution, emotion value group) P (x (k) | ej) for the j-th basic emotion in the n-dimensional internal state coordinate system is determined as shown in Equation 4 below.

そして、ω_ｊ，ｍ（ｋ）は、ｋ時間でｊ番目の感情分布のｍ番目のガウス分布またはガウスモード（Gaussian mode）の加重値であり、Ｍ_ｊ（ｋ）は、ｋ時間でｊ番目の感情分布が持っているガウスモードの数であり、μ_ｊ，ｍ（ｋ）は、ｋ時間でｊ番目の感情分布のｍ番目ガウスモードの中央値であり、Σ_ｊ，ｍ（ｋ）は、ｋ時間でｊ番目の感情分布のｍ番目ガウスモードの帯域幅であり、ｘ（ｋ）は、ｋ時間における感情状態入力値である。ガウスモードは、一つのガウス曲線を意味する。

Ω _{j, m} (k) is a weight value of the mth Gaussian distribution or Gaussian mode of the jth emotion distribution in k hours, and M _j (k) is jth in k hours. Μ _{j, m} (k) is the median of the mth Gaussian mode of the jth emotion distribution in k hours, and Σ _{j, m} (k) is , The bandwidth of the mth Gaussian mode of the jth emotion distribution at k hours, and x (k) is the emotional state input value at k hours. The Gaussian mode means one Gaussian curve.

(4) is a stochastic representation of the emotion value P _| probabilistic representation (probability value) for calculating the emotional value (e j _x (k)), the equation 5 by using the Bayes theorem (Bayes' rule) Is calculated as follows. Here, P (e _j | x (k)), which is the probability value of the j-th emotion, is used as a posterior probability.

ここで、Ｐ（ｘ（ｋ）｜ｅ_ｊ）は、尤度関数(likelihood function)として使われ、Ｐ（ｅ_ｊ）は、ｅ_ｊの事前確率（prior probability）であって、各感情値が選択される確率（感情値の確率的表現）であり、各感情値の事前確率の和は、１である。

Here, P (x (k) | e _j ) is used as a likelihood function, P (e _j ) is a prior probability of e _j , and each emotion value is The probability of being selected (probabilistic expression of emotion value), and the sum of the prior probabilities of each emotion value is 1.

According to the above formulas 3 to 5, when J basic emotions are given, a J-dimensional emotion vector e (k) having J P (e _j | x (k)) is determined. . e (k) is a composite emotion in which J emotion values are reflected as a probabilistic expression.

(5) Calculate Matching Function
Using the feedback information of the user centering on the vector point of the internal state coordinate system x , the emotional Gaussian mode (or Gaussian distribution) for learning is searched for as follows.

Here, the matching function (match _{j, m} (k + 1)) satisfies the following Expression 6.

ここで、matchｊ，ｍ（ｋ＋１）は、ｋ＋１時間でｊ番目の感情分布のうちｍ番目のガウス分布のマッチング関数を意味し、δは、マッチング範囲を示し（この時、δが大きければ、多くのガウスモード（ガウス分布）を学習するため、学習が速くなり得る）、σｊ，ｍ（ｋ）は、ｋ時間でｊ番目の感情分布のうちｍ番目のガウス分布の標準偏差を示し、ｘは、前記内部状態座標系の一点を示し、μｊ，ｍ（ｋ）は、ｋ時間でｊ番目の感情分布のうちｍ番目のガウス分布の中央値を示す。

Here, matchj, m (k + 1) means a matching function of the mth Gaussian distribution among the jth emotion distributions in k + 1 hours, and δ indicates a matching range (at this time, if δ is large, it is more Σj, m (k) indicates the standard deviation of the mth Gaussian distribution of the jth emotion distribution in k hours, and x is Represents one point of the internal state coordinate system , and μj, m (k) represents the median value of the mth Gaussian distribution of the jth emotion distribution in k hours.

  For reference, the matching function is a function defined by the Euclidean Distance calculation method. In addition to this, the Mahalanobis distance calculation method or the Minkowski distance calculation method is used. You may define a matching function.

(6) User Feedback Learning on Probability Distribution
After selecting the Gaussian mode as follows using the matching function, the Gaussian distribution is learned reflecting the user's intention.

  Learning is divided into two cases. The first case is a case where a Gaussian mode learned through a matching function is found (6-1), and the second case is a case where a Gaussian mode is not found (6-2). It is.

(6-1) ∃m∈ {1, 2,..., Mj (k)} such that matchj, m (k + 1) = 1 If any Gaussian distribution matched through the matching function can be found, Gaussian The distribution is learned. At this time, the selected, that is, matched Gaussian distribution and the non-selected, that is, unmatched Gaussian distribution are learned separately.

(6-1-1) m of each of matchj and m (k + 1) = 1
As a result, the mth Gaussian mode of the selected jth emotion distribution is learned. In this case, the three values of weight, median, and variance of the Gaussian distribution are learned.

In other words, when the matching function is 1, the m-th Gaussian distribution among the selected j-th emotion distribution is a weighted value (ωj, m (k + 1)) of the following Equation 7 using the feedback information. , The Gaussian distribution is changed to satisfy the median (μj, m (k + 1)) and the variance (σ2j, m (k + 1)).

ここで、ω_ｊ，ｍ（ｋ＋１）は、ｋ＋１時間でｊ番目の感情分布のうちｍ番目のガウス分布の加重値であり、λは、既定の基本加重値であり、ω_ｊ，ｍ（ｋ）は、ｋ時間でｊ番目の感情分布のうちｍ番目のガウス分布の加重値であり、μ_ｊ，ｍ（ｋ＋１）は、ｋ＋１時間でｊ番目の感情分布のうちｍ番目のガウス分布の中央値であり、ｘは、前記内部状態座標系の一点であり、μ_ｊ，ｍ（ｋ）は、ｋ時間でｊ番目の感情分布のうちｍ番目のガウス分布の中央値であり、σ^２ _ｊ，ｍ（ｋ＋１）は、ｋ＋１時間でｊ番目の感情分布のうちｍ番目のガウス分布の分散であり、σ^２ _ｊ，ｍ（ｋ）は、ｋ時間でｊ番目の感情分布のうちｍ番目のガウス分布の分散であり、［ ］^Ｔは、転置行列であり、

である。この時、Σ_ｊ，ｍ（ｋ）は、ｋ時間でｊ番目の感情分布のうちｍ番目のガウス分布の帯域幅である。

Here, ω _{j, m} (k + 1) is a weight value of the mth Gaussian distribution of the jth emotion distribution in k + 1 hours, λ is a default basic weight value, and ω _{j, m} (k ) Is a weighted value of the mth Gaussian distribution of the jth emotion distribution in k hours, and μ _{j, m} (k + 1) is the center of the mth Gaussian distribution of the jth emotion distribution in k + 1 hours. X is one point of the internal state coordinate system , μ _{j, m} (k) is the median value of the m-th Gaussian distribution of the j-th emotion distribution in k hours, and σ ² _{j , M} (k + 1) is the variance of the mth Gaussian distribution in the jth emotion distribution at k + 1 time, and σ ² _{j, m} (k) is the _mth of the jth emotion distribution at k time. The variance of the Gaussian distribution, [] ^T is the transpose matrix,

It is. At this time, Σ _{j, m} (k) is the bandwidth of the mth Gaussian distribution of the jth emotion distribution in k hours.

(6-1-2) match _{j, m} (k + 1) = 0 for each m
In addition, the mth Gaussian mode of the jth emotion distribution which is not selected is learned. In this case, only the weight value is learned.

In other words, when the matching function is 1, the m-th Gaussian distribution among the unselected j-th emotion distributions uses the feedback information, and the weight value (ω _{j, m} ( The Gaussian distribution is changed to satisfy k + 1)).

ここで、ω_ｊ，ｍ（ｋ＋１）は、ｋ＋１時間でｊ番目の感情分布のうちｍ番目のガウス分布の加重値であり、λは、既定の基本加重値であり、ω_ｊ，ｍ（ｋ）は、ｋ時間でｊ番目の感情分布のうちｍ番目のガウス分布の加重値である。

Here, ω _{j, m} (k + 1) is a weight value of the mth Gaussian distribution of the jth emotion distribution in k + 1 hours, λ is a default basic weight value, and ω _{j, m} (k ) Is a weighted value of the mth Gaussian distribution of the jth emotion distribution in k hours.

  Next, after the Gaussian distribution is changed, the emotion learning unit 400 normalizes weights of all Gaussian distributions corresponding to each emotion. That is, the sum of the weight values is set to 1.

_{(6-2) match j, m (} k + 1) = 0 for ∀m∈ {1,2, ..., M j (k)} On the other hand, when the, if not find a Gaussian mode to be matched, the feedback information Use to set up a new Gaussian mode. Therefore, the number M _j (k + 1) of the Gaussian modes possessed by the jth emotion distribution in k + 1 hours is M _j (k + 1) = M _j (k) +1.

Gauss other words, the emotion learning unit 400 (if there is no Gaussian distribution are matched) when the matching function is zero, which satisfies by using the feedback information, the weighted value of the following formula 9, the median and variance Generate a distribution.

  Then, after the Gaussian distribution is generated, the emotion learning unit 400 normalizes weights of all Gaussian distributions corresponding to each emotion. That is, the sum of the weight values is set to 1.

  Here, the median of the Gaussian distribution indicates what kind of emotion it affects, the variance means how widely the Gaussian distribution affects, and the weight is the value among the various Gaussian distributions for the same emotion It means how much influence the corresponding Gaussian distribution has. That is, if the weight value is high, it means that many corresponding values of the Gaussian distribution are reflected.

  Accordingly, since the Gaussian distribution at time k is different from the Gaussian distribution at time k + 1, another output value, that is, an emotion value is generated for the same input. Through this process, the emotion distribution is learned on the emotion space 930 as desired by the user. If this process is repeated, the robot can generate and express emotions as the user has trained.

  6A is a schematic diagram illustrating an emotion model according to an embodiment of the present invention, FIG. 6B is a schematic diagram illustrating an emotion model changed by learning, and FIGS. 6A and 6B illustrate psychological FFM ( A five-dimensional emotion space 930 is constructed based on the Five Factor Model. Therefore, as shown in FIG. 6A, the Gaussian distribution G1 before learning is changed to the Gaussian distribution G2 after learning by the user's education. Therefore, when it is assumed that the same emotion state input value is input, the emotion determination processing unit 200 determines different emotion values before learning (FIG. 6A) and after learning (FIG. 6B).

  On the other hand, as shown in FIG. 7, in the artificial emotion learning method according to the present invention, (a) the internal state calculation unit 100 maps the external input space 910 to the internal state space 920 and senses external environment information to detect the internal state. A step (S100) of generating an internal state at one point existing in the space 920 is included. That is, at this stage, an internal state vector value corresponding to one point of the internal state coordinate system is calculated.

  Next, (b) an emotion determination processing unit 200 provided with an emotion model in which emotion probability distributions for many types of emotions are formed on the internal state space 920, the input value of the internal state calculation unit 100 is used as the internal value. A step (S200) of determining an emotion value of the robot by matching with the state space 920 is included. The step (b) includes an emotion model in which an emotion probability distribution for J types of emotions is formed on the n-dimensional internal state coordinate system.

  Next, (c) the motion generation unit 300 includes a step (S300) of expressing an emotion with the emotion value determined by the emotion determination processing unit 200.

  Finally, (d) the emotion learning unit 400 receives the emotion value determined from the emotion determination processing unit 200 and the emotion type information input from the user input unit 500, and enters the emotion space 930. It includes a step (S400) of generating feedback information corresponding to an emotion value which is one existing point and changing the emotion probability distribution using the feedback information. That is, feedback information related to an emotion value corresponding to one point of the internal state coordinate system is generated, and the emotion probability distribution of the internal state coordinate system is changed using the feedback information.

  The present invention has been described with reference to the drawings and embodiments. However, the present invention is not limited to the specific embodiments, and those skilled in the art can make many modifications and variations without departing from the scope of the present invention. You will understand that there is. Moreover, the said drawings are shown in order to assist the understanding of the invention, and should not be understood as limiting the scope of the claims.

  The present invention is suitably used in a technical field related to an artificial emotion learning apparatus.

1000 Artificial Emotion Learning Device 100 Internal State Calculation Unit 200 Emotion Determination Processing Unit 300 Action Generation Unit 400 Emotion Learning Unit 500 User Input Unit

Claims (10)

An internal state calculation unit that detects external environment information and calculates an internal state vector value corresponding to one point of the internal state coordinate system;
With feelings probability distribution formed on the internal state coordinate system, generating a feeling of the robot emotion value of the internal state calculation section each emotion value in group indicated by the coordinate of the calculated the internal state vector values An emotion decision processing unit,
An action generation unit that determines an action of the robot and expresses an emotion by the emotion value determined by the emotion determination processing unit;
The emotion value determined from the emotion determination processing unit and the emotion type information input from the user input unit are input, and feedback information regarding the emotion value corresponding to one point of the internal state coordinate system is generated, An emotion learning unit that changes the emotion probability distribution of the internal state coordinate system using feedback information;
A user input unit for inputting emotion type information to the emotion learning unit;
Artificial emotion learning device.   2. The artificial emotion according to claim 1, wherein the emotion probability distribution is a Gaussian distribution having a weighted value, a median value, and a variance, or a Gaussian distribution obtained by combining a plurality of the Gaussian distributions. Learning device.   The artificial emotion learning apparatus according to claim 2, wherein the emotion learning unit searches for a Gaussian distribution corresponding to the feedback information using a matching function and changes the Gaussian distribution. 4. The artificial emotion learning apparatus according to claim 3, wherein the matching function (match _{j, m} (k + 1)) satisfies the following expression.

Here, match _{j, m} (k + 1) means a matching function of the mth Gaussian distribution among the jth emotion distributions in k + 1 hours, δ indicates a matching range, and σ _{j, m} (k) is , Indicates the standard deviation of the mth Gaussian distribution among the jth emotion distributions in k hours, x indicates one point of the internal state coordinate system , and μ _{j, m} (k) The median of the mth Gaussian distribution in the emotion distribution is shown. When at least one Gaussian distribution is 1 in the matching function, the m-th Gaussian distribution among the j-th emotion distributions for which the matching function is 1 uses the feedback information to calculate the next weight value (ω _{j , M} (k + 1)), median (μ _{j, m} (k + 1)), and variance (σ ² _{j, m} (k + 1)), the Gaussian distribution is changed. Artificial emotion learning device.

Here, ω _{j, m} (k + 1) is a weight value of the mth Gaussian distribution of the jth emotion distribution in k + 1 hours, λ is a default basic weight value, and ω _{j, m} (k ) Is a weighted value of the mth Gaussian distribution of the jth emotion distribution in k hours, and μ _{j, m} (k + 1) is the center of the mth Gaussian distribution of the jth emotion distribution in k + 1 hours. X is one point of the internal state coordinate system , μ _{j, m} (k) is the median value of the m-th Gaussian distribution of the j-th emotion distribution in k hours, and σ ² _{j , M} (k + 1) is the variance of the mth Gaussian distribution in the jth emotion distribution at k + 1 time, and σ ² _{j, m} (k) is the _mth of the jth emotion distribution at k time. The variance of the Gaussian distribution, [] ^T is the transpose matrix,

At this time, Σ _{j, m} (k) is the bandwidth of the m-th Gaussian distribution of the j-th emotion distribution in k hours. When at least one Gaussian distribution is 1 in the matching function, the m-th Gaussian distribution among the j-th emotion distributions for which the matching function is 0 uses the feedback information to obtain the following weight value (ω _j, The artificial emotion learning apparatus according to claim 4, wherein the Gaussian distribution is changed so as to satisfy _m (k + 1)).

Here, ω _{j, m} (k + 1) is a weight value of the mth Gaussian distribution of the jth emotion distribution in k + 1 hours, λ is a default basic weight value, and ω _{j, m} (k ) Is a weighted value of the mth Gaussian distribution of the jth emotion distribution in k hours.   The artificial emotion learning device according to claim 5 or 6, wherein the emotion learning unit normalizes weights of all Gaussian distributions corresponding to each emotion after the Gaussian distribution is changed. . When the matching function is 0 (when there is no matched Gaussian distribution), the emotion learning unit generates a Gaussian distribution that satisfies the following weight value, median, and variance using the feedback information. The artificial emotion learning apparatus according to claim 4.

  9. The artificial emotion learning apparatus according to claim 8, wherein the emotion learning unit normalizes weight values of all Gaussian distributions corresponding to each emotion after the Gaussian distribution is generated. An internal state calculating unit detecting external environment information, calculating an internal state vector value corresponding to one point of the internal state coordinate system, and generating an internal state;
An emotion determination processing unit provided with an emotion model in which an emotion probability distribution for j types of emotions is formed on an n-dimensional internal state coordinate system, the internal state vector value calculated by the internal state calculation unit as the internal state vector value Matching the state coordinate system to determine the emotional value of the robot;
A step of performing an emotion expression by determining an operation of the robot according to an emotion value determined by the emotion determination processing unit;
The emotion learning unit receives the emotion value determined from the emotion determination processing unit and the emotion type information input from the user input unit, and provides feedback information regarding the emotion value corresponding to one point of the internal state coordinate system. Generating and changing the emotion probability distribution of the internal state coordinate system using the feedback information;
Artificial emotion learning method including.

Images