Estimation device, information processing device, estimation method, and program

Abstract

To provide an estimation apparatus configured to realize both accuracy of estimation and followability of estimation for the change of state of an observation target, in estimating a state of the non-stationary observation target, an information processing apparatus, an estimation method, and a program.SOLUTION: An estimation apparatus 110 which estimates a state of an observation target includes: a data acquisition unit 121 which acquires observation data from the observation target; a first update unit 122 which updates contents of multiple hypotheses on the basis of the observation data; a second update unit 123 which updates indices indicating probabilities of the hypotheses on the basis of the observation data; and a state estimation unit 124 which outputs, as an estimated state, the most probable hypothesis out of the multiple hypotheses, on the basis of the observation data.SELECTED DRAWING: Figure 3

Description

[Technical Field]

[0001] The present invention relates to an estimation device, an information processing device, an estimation method, and a program. [Background technology]

[0002] Conventionally, there are known devices that estimate the state of an observed object based on observation data acquired from the observed object. For example, a device described in Patent Document 1 learns a model representing the state of the observed object, including transition probabilities between multiple states, based on time-series data of observation values ​​acquired from the observed object. Then, using data of observation values ​​prior to a predetermined time, the state of the observed object at a predetermined time is estimated based on the transition probabilities between multiple states included in the learned model. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Japanese Patent Application Laid-Open No. 2015-49726 Summary of the Invention [Problem to be solved by the invention]

[0004] Generally, when estimating the state of an observed object, the observation data obtained from the observed object contains uncertainty, and the estimated state of the observed object tends to vary each time observation data is obtained. Therefore, conventional devices estimate the state of the observed object while suppressing the impact of uncertainty in the observation data by using observation data over a long period of time.

[0005] However, when the state of the observed object is non-stationary and changes during observation, it is necessary to estimate the state of the observed object by using not only long-term observation data but also short-term observation data. Thus, when estimating the state of the observed object, if the observation data interval is long, the accuracy of the estimation increases but the ability to track changes in the state of the observed object decreases. Conversely, if the observation data interval is short, the ability to track changes in the state of the observed object increases but the accuracy of the estimation decreases. In other words, there is a problem in that a trade-off occurs between the accuracy of the estimation and the ability to track changes in the state of the observed object depending on the length of the observation data interval.

[0006] The present invention has been made in consideration of the above circumstances, and aims to provide an estimation device, an information processing device, an estimation method, and a program that, when estimating the state of a non-stationary observed object, can achieve both accuracy in estimation and responsiveness to changes in the state of the observed object. [Means for solving the problem]

[0007] An estimation device according to one aspect of the present invention is an estimation device that estimates the state of an observed object, and includes: a data acquisition unit that acquires observation data from the observed object; a first update unit that updates the content of each of a plurality of hypotheses based on the observation data; a second update unit that updates an index indicating the likelihood of each of the plurality of hypotheses based on the observation data; and a state estimation unit that outputs the most likely hypothesis of each of the plurality of hypotheses as an estimated state based on the observation data. [Effects of the Invention]

[0008] According to the present invention, when estimating the state of a non-stationary observation target, both accuracy of estimation and tracking ability can be achieved. [Brief explanation of the drawings]

[0009] [Figure 1] FIG. 2 is a diagram for explaining an outline of processing performed by an estimation device according to an embodiment. [Figure 2] FIG. 2 is a diagram illustrating an example of a processing flow of an estimation device according to an embodiment. [Figure 3] 1 is a block diagram showing a schematic configuration of an information processing device according to an embodiment; [Figure 4] FIG. 10 is a diagram illustrating an example of a time transition of an index indicating the likelihood of a hypothesis. [Figure 5] FIG. 10 is a diagram showing an example of the time transition of the content of a hypothesis. [Figure 6] FIG. 10 is a diagram illustrating an example of time-series data of the estimation results of the estimation device in response to changes in the state of the observed object. [Figure 7] FIG. 10 is a diagram showing an example of a process for forming the content of a hypothesis. [Figure 8] FIG. 10 is a diagram showing time-series data of the estimation results of the estimation device before and after a change in the state of the observation target. [Figure 9] FIG. 10 is a diagram illustrating the correlation between the responsiveness of an estimation result to a change in the state of an observation target and the accuracy of the estimation result. DETAILED DESCRIPTION OF THE INVENTION

[0010] Hereinafter, an embodiment of an estimation device, an information processing device, an estimation method, and a program according to the present invention will be described with reference to the drawings.

[0011] An estimation device according to this embodiment estimates the state of an observed object based on observation data acquired from the observed object. Specifically, as shown in FIG. 1 , the estimation device first defines multiple hypotheses. In this example, the hypotheses indicate the emotions of the observed object, with hypothesis h1 indicating that the object is happy, hypothesis h2 indicating that the object is angry, hypothesis h3 indicating that the object is sad, and hypothesis h4 indicating that the object is having fun. Next, the estimation device updates an index indicating the likelihood of each of the multiple hypotheses each time observation data is acquired from the observed object. The index indicating the likelihood of a hypothesis is, for example, the probability that the hypothesis will be established. The estimation device then estimates the most likely hypothesis from among the multiple hypotheses based on the observation data and outputs the estimated state of the observed object. Here, "estimating the most likely hypothesis from among the multiple hypotheses" means, for example, when there are n hypotheses, estimating the hypothesis with the highest probability of establishment from among the n hypotheses as the state of the observed object. In the example shown in FIG. 1, the estimation device estimates the subject's emotion from four candidate emotions: "joy," "anger," "sadness," and "happiness." In this case, the estimation device estimates the state of the observed object by tracking changes in the environment in a non-stationary environment. Here, a "non-stationary environment" refers to an environment in which the relationship between the state of the observed object and the observation data observed in that state fluctuates. For example, the estimation device sequentially estimates the subject's emotion, which changes over time. Furthermore, each time observation data is acquired from the observed object, the estimation device updates a model of the most likely hypothesis out of multiple hypotheses. Here, the "model" is, for example, a mathematical model that indicates the relationship between the state of the observed object and the observation data observed in that state. Using this model, the probability of acquiring observation data under conditions in which the hypothesis is valid is calculated as the content of the hypothesis.

[0012] FIG. 2 is a diagram illustrating an example of the flow of processing by the estimation device according to the present embodiment. In the example shown in the figure, the estimation device first calculates an index C(h i ) and the initial values ​​for each of the hypotheses C(d|h iSpecifically, the estimation device sets an initial value for the index C(h i ) as the initial value for P(h i ) and set the model C(d|h i ) as the initial value for P(d|h i ) to set the

[0013] Next, the estimation device calculates the observed data d at time t. t Then, the estimation device calculates the probability that each of the multiple hypotheses is true based on the following equation (1).

number

[0014] Next, the estimation device outputs the most likely hypothesis at time t+1 based on the following equation (2).

number

[0015] The estimation device also calculates the hypothesis h that has the highest probability of being true at time t based on the following equation (3): t max Update the model.

number

[0016] Thereafter, the process of acquiring observation data from the object of observation, updating the model of the hypothesis with the highest probability of being true based on the acquired observation data, and updating the index indicating the likelihood of each hypothesis is repeated.

[0017] FIG. 3 is a block diagram showing a schematic configuration of an information processing device according to this embodiment. As shown in the figure, the information processing device 100 includes, for example, a control unit 120 and a storage unit 130. The control unit 120 is realized by, for example, a hardware processor such as a CPU (Central Processing Unit) executing a program (software). Some or all of these components may be realized by hardware (including circuitry) such as an LSI (Large Scale Integration), an ASIC (Application Specific Integrated Circuit), an FPGA (Field-Programmable Gate Array), or a GPU (Graphics Processing Unit), or may be realized by a combination of software and hardware. The program may be stored in advance in a storage device such as a HDD or flash memory of the information processing device 100, or may be stored in a removable storage medium such as a DVD or CD-ROM, and installed in the HDD or flash memory of the information processing device 100 by inserting the storage medium into a drive device.

[0018] The control unit 120 includes, for example, a data acquisition unit 121, a first update unit 122, a second update unit 123, a state estimation unit 124, and an information output unit 125. In this embodiment, the estimation device 110 is configured with the data acquisition unit 121, the first update unit 122, the second update unit 123, the state estimation unit 124, and a storage unit 130.

[0019] The data acquisition unit 121 acquires observation data of the observation target from the outside. For example, the data acquisition unit 121 acquires the observation data of the observation target through an input device connected to the information processing device 100. The input device includes, for example, a camera, a microphone, etc.

[0020] The first updating unit 122 updates an index indicating the likelihood of each of the plurality of hypotheses based on the observation data acquired by the data acquiring unit 121. The first updating unit 122, for example, calculates the probability that each of the plurality of hypotheses will be established as an index indicating the likelihood of each of the plurality of hypotheses. The first updating unit 122, for example, reads out an index indicating the likelihood of each of the plurality of hypotheses before updating from the storage unit 130 as pre-update index data 132, and stores an index indicating the likelihood of each of the plurality of hypotheses after updating in the storage unit 130 as post-update index data 132 based on the pre-update index data 132 and the observation data acquired by the data acquiring unit 121.

[0021] The second update unit 123 performs the following update based on the observation data acquired by the data acquisition unit 121: The second updating unit 123 updates a model of the most likely hypothesis among the multiple hypotheses. The most likely hypothesis is, for example, the hypothesis with the highest probability of being true, calculated as an index indicating the likelihood of the hypothesis, among the multiple hypotheses. The second updating unit 123 stores the model of the most likely hypothesis in the storage unit 130 as updated hypothesis data 131 for the most likely hypothesis, based on, for example, pre-update hypothesis data 131 for the most likely hypothesis and the observation data acquired by the data acquiring unit 121.

[0022] The state estimation unit 124 outputs the most likely hypothesis among the multiple hypotheses as the estimated state of the observed object based on the observation data acquired by the data acquisition unit 121. The state estimation unit 124, for example, compares the probability that each of the multiple hypotheses calculated by the first update unit 122 will be established, and estimates the hypothesis with the highest probability as the most likely hypothesis.

[0023] The information output unit 125 outputs information indicated by the most likely hypothesis from among the multiple hypotheses estimated by the state estimation unit 124. The information output unit 125 outputs, for example, the information indicated by the most likely hypothesis through an output device connected to the information processing device 100. The output device includes, for example, a display, a speaker, etc.

[0024] FIG. 4 is a diagram showing an example of the time transition of an index indicating the likelihood of a hypothesis. In the example shown, 11 hypotheses (h1 to h 11 ) over time. As shown in the figure, the probability that each hypothesis is true fluctuates in accordance with changes in the state of the observed object. Specifically, when the state of the observed object changes, the probability that the hypothesis corresponding to the state of the observed object after the change is true increases, while the probability that the hypothesis corresponding to the state of the observed object before the change is true decreases. Furthermore, the probability that other hypotheses are true also fluctuates in accordance with changes in the state of the observed object. Fluctuations in the probability that each of the above hypotheses is true occur every time the state of the observed object changes.

[0025] FIG. 5 is a diagram showing an example of the time transition of the content of hypotheses. In the example shown, 11 hypotheses (h1 to h 11 ) is true. As shown in the figure, the probability of obtaining observational data under the conditions in which each hypothesis is true varies in accordance with changes in the state of the observed object. In other words, when the state of the observed object changes, the observational data obtained from the observed object changes, and the probability of obtaining observational data under the conditions in which each hypothesis is true also varies. Specifically, when the state of the observed object changes, the probability of obtaining observational data under the conditions in which the hypothesis corresponding to the state of the observed object after the change is true increases, while the probability of obtaining observational data under the conditions in which the hypothesis corresponding to the state of the observed object before the change is true decreases. Furthermore, as the state of the observed object changes, the probability of obtaining observational data under the conditions in which the other hypotheses are true also varies. The fluctuations in the probability of obtaining observational data under the conditions in which each of the above hypotheses is true occur every time the state of the observed object changes.

[0026] FIG. 6 is a diagram showing an example of time-series data of the estimation results of an estimation device in response to a change in the state of an observed object. In the example shown in the figure, the actual change in the state of the observed object is designated as the "correct answer," and the probability that a hypothesis corresponding to the state of the observed object after the change is established is designated as the "example." In this example, when the number of trials is relatively small, there is a discrepancy between the "correct answer" and the "example." However, as the number of trials increases, the discrepancy between the "correct answer" and the "example" tends to become smaller. In other words, as the number of trials increases, the accuracy of the estimation results by the estimation device according to this embodiment gradually increases.

[0027] FIG. 7 is a diagram showing an example of the process of forming the contents of hypotheses. In the example shown in the figure, 11 hypotheses (h1 to h 11 ) shows the process of forming the content of hypotheses for each of the multiple hypotheses. In this example, as the number of trials increases, the content of each hypothesis is formed in a different manner. That is, each time observation data is acquired from the observed object, the index indicating the likelihood of each of the multiple hypotheses and the content of each of the multiple hypotheses are updated, and the content of each of the multiple hypotheses is formed to reflect the tendency of changes in the state of the observed object.

[0028] FIG. 8 is a diagram showing time-series data of the estimation results of the estimation device before and after a change in the state of the observed object. In the example shown in the figure, the actual change in the state of the observed object is designated as the "correct answer," the time-series data of the estimation results by the estimation device 110 according to the present embodiment is designated as the "Example," the time-series data of the estimation results by the estimation device of the first comparative example is designated as the "Comparative Example 1," and the time-series data of the estimation results by the estimation device of the second comparative example is designated as the "Comparative Example 2." In this example, the estimation device of the first comparative example updates the content of each of multiple hypotheses based on observation data acquired from the observed object and estimates a change in the state of the observed object by calculating a time average of the time-series data of the estimation results over a first interval length. Meanwhile, the estimation device of the second comparative example updates the content of each of multiple hypotheses based on observation data acquired from the observed object and estimates a change in the state of the observed object by calculating a time average of the time-series data of the estimation results over a second interval length longer than the first interval length. As shown in the figure, the interval length for calculating the time average is shorter in "Comparative Example 1" than in "Comparative Example 2," and therefore the estimation result has high tracking ability to changes in the state of the observed object. Furthermore, the estimation result of "Example" has even higher tracking ability to changes in the state of the observed object than in "Comparative Example 1."

[0029] FIG. 9 illustrates the correlation between the tracking ability and accuracy of the estimation results of an estimation device relative to changes in the state of the observed object. In the example shown in the figure, the characteristics of the estimation results by the estimation device 110 according to the present embodiment are shown as an “Example,” and the characteristics of the estimation results by estimation devices that update the content of each of multiple hypotheses based on observation data acquired from the observed object when the interval length for calculating the time average is continuously changed, including the estimation device of the first comparative example and the estimation device of the second comparative example, are shown as “Comparative Example.” As shown in the figure, in the “Comparative Example,” as the interval length for calculating the time average is increased, the influence of errors in the time-series data is more easily offset, resulting in higher accuracy of the estimation results, but lower tracking ability of the estimation results relative to changes in the state of the observed object. In other words, there is a trade-off between the tracking ability of the estimation results relative to changes in the state of the observed object and the accuracy of the estimation results. In contrast, the "Example" not only updates the content of each of the multiple hypotheses based on the observation data obtained from the observed object, but also updates the index indicating the likelihood of each of the multiple hypotheses, thereby improving both the responsiveness and accuracy of the estimation results to changes in the state of the observed object compared to the "Comparative Example," thereby eliminating this trade-off.

[0030] As described above, according to the above embodiment, the following effects can be obtained. (1) The estimation device 110 updates the content of each of the multiple hypotheses based on observation data acquired from the observed object, updates an index indicating the likelihood of each of the multiple hypotheses, and outputs the most likely hypothesis of the multiple hypotheses as the estimated state based on the observation data acquired from the observed object. That is, when observation data is acquired from the observed object, not only is the index indicating the likelihood of each of the multiple hypotheses updated, but the content of each of the multiple hypotheses is also updated. This makes it possible to achieve both accuracy in estimation and the ability to track changes in the state of the observed object when estimating the state of a non-stationary observed object.

[0031] (2) The index indicating the likelihood of a hypothesis is the probability that the hypothesis is true, the content of the hypothesis is the probability that observed data will be obtained under conditions under which the hypothesis is true, and the most likely hypothesis is the hypothesis with the highest probability of being true. This allows the content of each of the multiple hypotheses and the index indicating the likelihood of each of the multiple hypotheses to be defined as continuous values, making it possible to quantitatively evaluate the most likely hypothesis among the multiple hypotheses.

[0032] (3) The estimation device 110 calculates the probability that each hypothesis is true under the conditions under which the observation data is acquired, for each of the multiple hypotheses, and updates the probability that each hypothesis is true based on the calculated probability. That is, when observation data is acquired from the observation target, the content of the hypothesis is updated, and an index indicating the likelihood of each of the multiple hypotheses is updated based on the content of the updated hypothesis. This allows the content of the hypothesis updated in response to the acquisition of observation data to be reflected in the index indicating the likelihood of each of the multiple hypotheses.

[0033] (4) Each time observation data is acquired from the observed object, the estimation device 110 updates the content of each of the multiple hypotheses and updates the index indicating the likelihood of each of the multiple hypotheses. That is, as time passes since the state of the observed object changes, the content of each of the multiple hypotheses and the index indicating the likelihood of each of the multiple hypotheses are repeatedly updated to reflect the state of the observed object after the change. This makes it possible to achieve both higher levels of estimation accuracy and tracking of changes in the state of the observed object when estimating the state of a non-stationary observed object.

[0034] (5) The hypotheses indicate the emotions of the subject, and the estimation device 110 acquires data about the subject and, based on the acquired data about the subject, estimates the emotion of the subject from among candidate emotions of the subject indicated by each of the multiple hypotheses. This makes it possible to achieve both accuracy in estimation and responsiveness to changes in the state of the observed subject when estimating the emotion of the subject by tracking changes in the emotion of the observed subject.

[0035] (6) The information processing device 100 outputs information indicated by the most likely hypothesis among the multiple hypotheses estimated by the estimation device 110. This allows the user to be notified of changes in the state of the non-stationary observed object estimated by the estimation device 110 when estimating the state of the observed object.

[0036] As described above, according to the estimation device 110 of this embodiment, estimation of the state of the observed object and updating (learning) of the hypothesis content are simultaneously executed. Furthermore, as shown in FIG. 2, the ratio of estimation processing to learning processing is automatically determined by the estimation device 110, and both processing can be executed while maintaining a balance between them. Conventionally, there have been known methods for updating a learning model by alternately executing estimation processing and learning processing, thereby improving estimation accuracy, but there has been no method for always executing estimation processing and learning processing simultaneously, as in the present invention. Furthermore, the weights (parameters) of the estimation processing and learning processing are also automatically determined, so the user does not need to decide.

[0037] (Application example) An example of application of the estimation device according to this embodiment will be described below. As mentioned above, the estimation device according to this embodiment can be applied to human emotion estimation. Specifically, for example, it can be applied to estimating a person's emotions (joy, anger, sadness, happiness, etc.) from the tone (volume, tempo, pitch, etc.) of the person's speech. In emotion estimation, in addition to noise in the estimation result, the emotions themselves fluctuate (non-steady state). Therefore, by applying the estimation device of the present invention, it is possible to achieve both accuracy and tracking of the estimation.

[0038] Furthermore, the estimation device according to this embodiment can be applied to general machine learning tasks by treating the above formulas (1) to (3) as multidimensional normal distributions. Hereinafter, online advertising, recommendation systems, and robot behavior control will be described as specific examples of applications to reinforcement learning tasks.

[0039] (Online advertising) The estimation device according to this embodiment can be applied to learning a model that presents online advertisements suitable for a user. When applied to estimating suitable online advertisements, for example, advertisements for products preferred by a user are estimated and presented based on the user's past browsing history and purchase history. Because user preferences are constantly changing, it is necessary to estimate non-stationary states. The estimation device acquires user behavior regarding the presented advertisements (whether or not the user ended up purchasing the product, etc.) as observation data, and performs machine learning of the model (updating the content of hypotheses and indicators) based on the observation data. In this way, the estimation process for online advertisements suitable for a user and reinforcement learning of the model can be performed simultaneously.

[0040] (Recommendation system) The estimation device according to this embodiment can be applied to learning a model that presents recommended products to a user. When applied to estimating recommended products, for example, products that a user prefers are estimated and presented based on the user's past browsing history and purchase history. Because user preferences are constantly changing, it is necessary to estimate non-stationary states. The estimation device acquires the user's behavior regarding the presented products (whether or not they ended up purchasing, etc.) as observation data, and performs machine learning of the model (updating the content of hypotheses and indicators) based on the observation data. In this way, it is possible to simultaneously perform the process of estimating recommended products for a user and the reinforcement learning of the model.

[0041] (Robot behavior control) The estimation device according to this embodiment can be applied to learning a model that selects an action that a communication robot will take toward a person it is facing. Specifically, it estimates the action that will most satisfy the user from among multiple patterns of actions that the robot can take toward a person. Since the action desired by the user is likely to change even while the robot is searching for an action to take, it is necessary to estimate non-stationary states. The estimation device acquires the user's actual reaction (voice, facial expression, etc.) to the selected action as observation data, and performs machine learning of the model (updating the content of hypotheses and indicators) based on the observation data. In this way, by simultaneously performing the process of estimating the action to be taken and the reinforcement learning of the model, it is possible to control the robot so that it can immediately respond to changing user requests.

[0042] (Feedback to the learning device) By combining the estimation results output by the estimation device according to this embodiment with actual observation data, highly reliable training data for machine learning can be created. In other words, by using data that has been labeled based on observation data for hypotheses of estimation results as training data, the generalization performance of the learning model can be improved.

[0043] The above-described embodiments are intended to facilitate understanding of the present invention and are not intended to limit the present invention. The elements of the embodiments, as well as their arrangement, materials, conditions, shapes, sizes, etc., are not limited to those illustrated and can be modified as appropriate. Furthermore, configurations shown in different embodiments can be partially substituted or combined with each other. [Explanation of symbols]

[0044] 100...information processing device, 110...estimation device, 120...control unit, 121...data acquisition unit, 122...first update unit, 123...second update unit, 124...state estimation unit, 125...information output unit, 130...storage unit, 131...hypothesis data, 132...index data.

Claims

1. An estimation device that estimates a state of an observation object, a data acquisition unit that acquires observation data from the observation target; a first update unit that updates the probability that each of a plurality of hypotheses regarding the state of the observation target is established based on the observation data; a second update unit that updates a model of a hypothesis that is most likely to be true based on the observation data; a state estimation unit that outputs a hypothesis that is most likely to be established among the plurality of hypotheses as a state of the observation target; The model is a mathematical model that calculates the probability that the observation data will be acquired under conditions in which a hypothesis holds.

2. Each time the data acquisition unit acquires the observation data, The estimation device according to claim 1 , wherein the first update unit updates the probability that the hypothesis is true, and the second update unit updates the model simultaneously.

3. The hypothesis describes the subject's emotions, the data acquisition unit acquires data related to the subject; 3. The estimation device according to claim 1, wherein the state estimation unit estimates an emotion of the subject from among candidate emotions of the subject indicated by each of the plurality of hypotheses, based on the data about the subject acquired by the data acquisition unit.

4. The estimation device according to any one of claims 1 to 3; an information output unit that outputs information indicated by the hypothesis that is most likely to be true, as estimated by the estimation device.

5. A method for a computer to estimate a state of an observed object, comprising: a data acquisition step in which a computer acquires observation data from the observation target; a first updating step in which the computer updates the probability that each of a plurality of hypotheses regarding the state of the observed object is established based on the observation data; a second updating step in which the computer updates the model of the hypothesis that is most likely to be true based on the observation data; a state estimation step of outputting, by the computer, a hypothesis that is most likely to be established from among the plurality of hypotheses as a state of the observed object; An estimation method, wherein the model is a mathematical model that calculates the probability that the observed data will be obtained under conditions where a hypothesis holds.

6. On the computer, a data acquisition process for acquiring observation data from the observation target; a first update process for updating the probability that each of a plurality of hypotheses regarding the state of the observation target is established based on the observation data; a second update process for updating the model of the hypothesis that is most likely to be true based on the observation data; a state estimation process for outputting a hypothesis with the highest probability of being true from among the plurality of hypotheses as a state of the observed object; The program, wherein the model is a mathematical model that calculates the probability that the observation data will be obtained under conditions in which a hypothesis holds.

Images