Action prediction device and program

Abstract

An action prediction device according to one aspect comprises a prediction unit that predicts a time series of progress achieved by an agent on the basis of an action model in which the cost of the most recent action is amplified. The prediction unit is configured to predict the time series of the progress on the basis of a cost function in which the cost does not change for progress that is below a threshold value, and the cost increases in accordance with the progress for progress at or above the threshold value.

Description

Behavior prediction device and program

[0001] The embodiments relate to a behavior prediction device and a program.

[0002] Present bias is known as a model for predicting human behavior as people strive to achieve their goals. Present bias is a model that overvalues ​​current gains and losses and undervalues ​​future gains and losses. Modeling human behavior with present bias in mind is expected to lead to highly accurate predictions and optimal interventions.

[0003] For example, Non-Patent Document 1 discloses a model of human behavior under the influence of present bias.

[0004] Yasunori Akagi, Naoki Marumo, Takeshi Kurashima, “Analytically Tractable Modes for Decision Making under Present Bias”, arXiv: 2302.08132v1 [cs.GT], February, 2023.David Laibson, “Golden Eggs and Hyperbolic Discounting”, The Quarterly Journal of Economics, May, 1997.

[0005] Modeling human behavior with present bias requires modeling the cost of making progress, but predicting human behavior with high accuracy requires realistic estimates of the cost of making progress.

[0006] For example, when considering a task of achieving a step count goal within a certain period of time, it is reasonable to think that the steps inevitably incurred during commuting, etc., are free of cost, while trying to earn additional steps incurs a cost. In this way, in reality, it is expected that there will be no cost until a certain level of progress is made, and only once a certain level of progress is reached will a cost be incurred. In other words, there is room for improvement in modeling realistic human behavior while taking present bias into account.

[0007] The present invention has been made in light of the above circumstances, and its purpose is to provide a behavior prediction device and program that model realistic human behavior while taking present bias into account.

[0008] A behavior prediction device according to one aspect includes a prediction unit configured to predict a time series of progress made by an agent based on a behavior model in which the cost of a most recent action is amplified, the prediction unit being configured to predict the time series of the progress based on a cost function in which the cost remains unchanged for progress below a threshold and the cost increases in accordance with progress equal to or greater than the threshold.

[0009] According to the embodiment, it is possible to provide a behavior prediction device and a program that model realistic human behavior while taking present bias into consideration.

[0010] Fig. 1 is a block diagram showing an example of the hardware configuration of a behavior prediction device according to an embodiment. Fig. 2 is a block diagram showing an example of the functional configuration of the behavior prediction device according to an embodiment. Fig. 3 is a diagram showing an example of the data structure of model information stored in the behavior prediction device according to an embodiment. Fig. 4 is a diagram showing an example of the data structure of prediction results stored in the behavior prediction device according to an embodiment. Fig. 5 is a flowchart showing an example of behavior prediction processing in the behavior prediction device according to an embodiment.

[0011] Hereinafter, embodiments will be described with reference to the drawings. In the following description, components having the same functions and configurations will be given the same reference numerals.

[0012] 1. Configuration 1.1 Hardware Configuration First, the hardware configuration of the behavior prediction device according to the embodiment will be described.

[0013] FIG. 1 is a block diagram illustrating an example of a hardware configuration of a behavior prediction device according to an embodiment.

[0014] As shown in FIG. 1 , the behavior prediction device 1 includes, for example, a control circuit 11 , a storage 12 , a communication module 13 , an interface 14 , a drive 15 , and a storage medium 16 .

[0015] The control circuit 11 is a circuit that controls the overall components of the behavior prediction device 1. The control circuit 11 includes a CPU (central processing unit), a RAM (random access memory), a ROM (read only memory), etc. The CPU of the control circuit 11 controls the entire behavior prediction device 1 in accordance with a program stored in the ROM of the control circuit 11. The RAM of the control circuit 11 has a working area for the CPU of the control circuit 11. The ROM of the control circuit 11 stores programs and the like used by the behavior prediction device 1.

[0016] The storage 12 includes, for example, a hard disk drive (HDD) or a solid state drive (SSD). The storage 12 stores information used in various processes in the behavior prediction device 1.

[0017] The communication module 13 is a circuit used for transmitting and receiving data between the behavior prediction device 1 and an external device. The communication module 13 may be configured to connect the behavior prediction device 1 to a network (not shown).

[0018] The interface 14 is an interface that manages communication between a user and the control circuit 11. The interface 14 includes input devices and output devices. The input devices include, for example, a keyboard, a touch panel, and operation buttons. The output devices include, for example, an LCD (Liquid Crystal Display) or an EL (Electroluminescence) display, a printer, etc. The interface 14 converts input from the user into an electrical signal and then transmits it to the control circuit 11. The interface 14 outputs the execution result based on the user input to the user.

[0019] The drive 15 is a device for reading software stored in the storage medium 16. The drive 15 includes, for example, a CD (Compact Disk) drive or a DVD (Digital Versatile Disk) drive.

[0020] The storage medium 16 is a medium that stores software electrically, magnetically, optically, mechanically, or chemically. The storage medium 16 may store a program used by the behavior prediction device 1.

[0021] 1.2 Functional Configuration Next, the functional configuration of the behavior prediction device according to the embodiment will be described.

[0022] FIG. 2 is a block diagram illustrating an example of a functional configuration of the behavior prediction device according to the embodiment.

[0023] 2 , the control circuit 11 of the behavior prediction device 1 functions as a computer including an input unit 21, a prediction unit 22, and an output unit 23. The storage 12 of the behavior prediction device 1 stores model information 24 and a prediction result 25.

[0024] The input unit 21 is a functional block for acquiring model information 24 from a user via the interface 14. The model information 24 is information used when modeling human behavior. Details of the model information 24 will be described later. The input unit 21 stores the input model information 24 in the storage 12.

[0025] The prediction unit 22 is a functional block that executes a behavior prediction process and predicts human behavior while taking into account present bias. The prediction unit 22 constructs a behavior model that takes into account the present bias of the human being to be predicted, based on model information 24 stored in the storage 12. When constructing the behavior model, the prediction unit 22 applies a cost function that can express a state in which no cost is incurred until a certain amount of progress is achieved and only when a certain amount of progress is achieved, when calculating the cost for generating progress. This allows the prediction unit 22 to construct a more realistic human behavior model. The prediction unit 22 predicts human behavior based on the constructed behavior model. The prediction unit 22 stores a prediction result 25 in the storage 12. Details of the behavior model, the cost function, and the prediction result 25 will be described later.

[0026] The output unit 23 is a functional block for presenting the prediction result 25 to the user via the interface 14. The output unit 23 reads out and outputs the prediction result 25 stored in the storage 12. The output format is arbitrary, and examples include display on a display, output on a printer, etc.

[0027] 1.3 Behavior Model A behavior model applied to the behavior prediction process in the behavior prediction device according to the embodiment will be described.

[0028] 1.3.1 Behavioral Model First, we will explain the human behavior model. Hereinafter, a human who acts to achieve a certain goal will also be called an "agent."

[0029] The agent acts to achieve a target progress θ within a time limit T. If the target progress θ is achieved within the time limit T, the agent is given a reward R. Here, the time limit T is a positive integer. The target progress θ is a real number greater than or equal to 0. The reward R is a real number greater than or equal to 0.

[0030] The state of an agent is expressed as (t, x), where time t ∈ {0, 1, ..., T} is the index of time and x, a real number equal to or greater than 0, is the progress up to the corresponding time t. The agent progresses from the initial state (0, 0) to (1, x １ ), (2, x ２ ), ..., (T, x Ｔ ) is a sequence of states. ｔ ) to state (t+1, x ｔ＋１ ) has a cost. ｔ ) to state (t+1, x ｔ＋１ The cost incurred in transitioning to c(x) is calculated using cost function c. ｔ＋１ -x ｔ ) is expressed as

[0031] The agent repeats the process of "selecting the sequence that minimizes the cost from all sequences of states that it can take in the future, and taking one step along that sequence" within the time limit T. Specifically, the agent selects the sequence that minimizes the cost from all sequences of states that it can take in the future, and takes one step along that sequence. ｔ－１ ) is the next state sequence (t, yｔ ), ..., (T, y Ｔ ) Cost CT when following ｔ－１ is evaluated as shown in the following formula (1).

[0032] The first term of equation (1) is the state (t-1, x ｔ－１ ) to the state (t, y ｔ ) where the parameter β is a real number greater than 0 and less than or equal to 1. The first term in equation (1) is amplified by a coefficient 1 / β≧1 compared to the transition cost in the subsequent step. As such, the parameter β is also called the present bias parameter because it is used to express the agent's tendency to overestimate the most recent cost. The formulation of the effect of present bias using the parameter β is based on a discounting method called quasi-hyperbolic discounting.

[0033] The second term of equation (1) is the state (t, y ｔ ) to the state (T, y Ｔ ) The second term corresponds to the sum of the transition costs up to the last action. Unlike the first term, the second term does not correspond to the most recent action, so there is no effect of amplification of the transition cost by the parameter β.

[0034] The third term in equation (1) corresponds to the reward. That is, the third term is the final progress y Ｔ If y is equal to or greater than the target progress θ, the agent will receive a reward R. Ｔ If ≥ θ, then 1[y Ｔ ≧θ]=1, and y Ｔ <θ, 1[y Ｔ ≧θ]=0.

[0035] The agent is in the state (t-1, x ｔ－１ ) to the next state, the cost CT ｔ－１ The sequence of states (t, y ｔ ＊ ), ..., (T, y Ｔ ＊ ) and calculate the state (t-1, x ｔ－１ ) to the state (t, y ｔ ＊ However, if there are multiple sequences with the same cost, the agent will transition to the progress yｔ Mathematically, the sequence of states is x ０ :=0 and t=1, . . . , T, it is calculated by the following equation (2).

[0036] 1.3.2 Cost Function Next, the cost function will be explained.

[0037] The cost function c(Δ) for obtaining the cost required to obtain the progress Δ is modeled as shown in the following equation (3).

[0038] Here, the parameter α is a real number greater than 1. The offset w is a real number greater than or equal to 0. The cost function c(Δ) shown in equation (1) can express a state in which no cost is incurred until progress equal to or greater than the offset w is made, and only when progress equal to or greater than the offset w is made does a cost begin to be incurred. Furthermore, the cost function c(Δ) shown in equation (1) expresses that progress never decreases.

[0039] 1.3.3 Analytical prediction of the state Next, the state (t, x ｔ ) analytical predictions are explained.

[0040] The prediction unit 22 predicts the progress x at time t based on the following theorems (a) and (b). ｔ (a): "If wT≧θ, then x ｔ = wt" (b): "If wT < θ, However, p ｔ teeth and t ＊ teeth If there is no such t, ​​then t ＊ = T. ” Below, by theorems (a) and (b), the progress x ｔ We present a proof that can be analytically predicted.

[0041] First, regarding theorem (a), if wT≧θ, then x ｔ = wt minimizes the transition cost at each time and minimizes the progress y ｔ＋１ , ..., y ＴIt is easy to verify that maximizes

[0042] Next, for theorem (b), it is sufficient to prove that the following recurrence formula, formula (4), holds.

[0043] Equation (2) can be rearranged to become the following equation (5). Note that Jensen's inequality is used in the final transformation of the equation.

[0044] Furthermore, by rearranging a part of equation (5), it can be transformed into the following equation (6). However, in the final transformation of the equation, the following equation (7), which is an inequality derived from Hölder's inequality, was used.

[0045] The condition for the equality to be satisfied is as shown in the following equation (8).

[0046] By rearranging the above equation (8), the following equation (9) is obtained.

[0047] Therefore, the following equation (10) is obtained.

[0048] From the above, the state x one step before ｔ－１ If satisfies the following formula (11), then x ｔ =(θ−wT)+wt+p ｔ [x ｔ－１ -(θ-wT)-w(t-1)], otherwise x ｔ = x ｔ－１ +w. That is, it is shown that the formula (4) holds.

[0049] From the above, it can be seen that the prediction unit 22 can analytically solve the agent behavior model based on the cost function c(Δ) that takes the offset w into account by performing calculations in accordance with theorems (a) and (b).

[0050] 1.3.4 Model Information FIG. 3 is a diagram illustrating an example of the data structure of model information stored in the behavior prediction device according to the embodiment.

[0051] 3, the model information 24 includes parameters α and β, an offset w, a time limit T, a target progress θ, and a reward R, which are necessary for the prediction unit 22 to predict a state sequence based on the above-described behavioral model. The example in FIG. 3 shows a case where the time limit T is 20, the parameter β is 0.6, the reward R is 100, the offset w is 20, the parameter α is 3, and the target progress θ is 40.

[0052] Note that FIG. 3 is merely an example, and each of the parameters α and β, the offset w, the time limit T, the target progress θ, and the reward R can take any value within the above-mentioned domain of definition.

[0053] 1.3.5 Prediction Results FIG. 4 is a diagram illustrating an example of a data structure of prediction results stored in the behavior prediction device according to the embodiment.

[0054] As shown in FIG. 4, the prediction result 25 is a sequence of states (0, x) analytically predicted by the prediction unit 22 based on the above-described behavioral model. ０ ), (1, x １ ), ...(T, x Ｔ 4 shows a case where the progress is 0 at the initial time t=0, the progress is 10 at time t=1 one step later, and the progress finally reaches 40 at time T.

[0055] 1.2 Operation Next, the operation of the behavior prediction device according to the embodiment will be described.

[0056] FIG. 5 is a flowchart illustrating an example of a behavior prediction process in the behavior prediction device according to the embodiment.

[0057] 5 , based on a user's operation of the interface 14 (start), the input unit 21 inputs a time limit T, a target progress θ, a reward R, an offset w, and parameters α and β (S11). The input unit 21 stores the input information in the storage 12 as model information 24.

[0058] The prediction unit 22 predicts the sequence of the agent's states based on the model information 24 input in the process of S11 (S12).

[0059] Specifically, the prediction unit 22 compares wT with θ to determine which is larger.

[0060] If wT≧θ, the prediction unit 22 predicts the sequence using theorem (a). In this case, the prediction unit 22 predicts the progress x at all times t between 0 and T. ｔ For x ｔ = wt is applied.

[0061] If wT<θ, the predictor 22 predicts the sequence using theorem (b). In this case, the predictor 22 predicts the sequence from the state x ｔ－１ It is further determined whether satisfies the formula (11).

[0062] State one step before x ｔ－１ satisfies the formula (11), the prediction unit 22 ｔ =(θ−wT)+wt+p ｔ [x ｔ－１ -(θ-wT)-w(t-1)] is applied. On the other hand, the state x ｔ－１ does not satisfy the formula (11), the prediction unit 22 ｔ = x ｔ－１ Apply +w.

[0063] By performing the above-described processing, the prediction unit 22 predicts the sequence of states (0, x ０ ), ..., (T, x Ｔ The prediction unit 22 stores the generated sequence as a prediction result 25 in the storage 12 .

[0064] The output unit 23 outputs the sequence predicted in the process of S12 to the user via the interface 14 (S13).

[0065] When the process of S13 ends, the behavior prediction process ends (END).

[0066] 1.3 Effects of the Embodiment According to the embodiment, the prediction unit 22 predicts the time series of progress made by an agent based on a behavioral model in which the cost of the most recent action is amplified. The prediction unit 22 predicts the time series of progress based on a cost function c(Δ) in which the cost remains unchanged for progress less than the offset w, and increases in accordance with progress for progress equal to or greater than the offset w. Specifically, the cost function c(Δ) sets the cost to 0 when the progress is equal to or greater than 0 and less than the offset w, and sets the cost to (Δ-w) when the progress is equal to or greater than the offset. α This makes it possible to express a state in which no cost is incurred until progress equal to or greater than the offset w is made, and only when progress equal to or greater than the offset w is made does a cost begin to be incurred. This makes it possible to analytically predict human behavior that is more realistic than a behavioral model based on the cost function c(Δ) that does not consider the offset w. Furthermore, by predicting human behavior using a cost function that is more realistic in this way, it becomes possible to solve the intervention timing optimization problem with higher accuracy.

[0067] 3. Modifications, etc. Various modifications can be applied to the above-described embodiment.

[0068] In the above-described embodiment, the case where the program that executes the behavior prediction process is executed by the behavior prediction device 1 has been described, but this is not limiting. For example, the program that executes the behavior prediction process may be executed by a computing resource on the cloud.

[0069] The present invention is not limited to the above-described embodiments, and various modifications can be made in the implementation stage without departing from the spirit of the invention. Furthermore, the embodiments may be implemented in appropriate combinations, in which case the combined effects can be obtained. Furthermore, the above-described embodiments include various inventions, and various inventions can be extracted by combining selected elements from the disclosed elements. For example, if the problem can be solved and the desired effect can be obtained even if some elements are deleted from all elements shown in the embodiments, the configuration from which these elements are deleted can be extracted as an invention.

[0070] REFERENCE SIGNS LIST 1... Behavior prediction device 11... Control circuit 12... Storage 13... Communication module 14... Interface 15... Drive 16... Storage medium 21... Input unit 22... Prediction unit 23... Output unit 24... Model information 25... Prediction result

Claims

1. A behavior prediction device comprising: a prediction unit that predicts a time series of progress made by an agent based on a behavioral model in which the cost of the most recent action is amplified; and the prediction unit is configured to predict the time series of progress based on a cost function in which cost remains unchanged for progress below a threshold, and cost increases in accordance with progress above the threshold.

2. The behavior prediction device according to claim 1, wherein the prediction unit is configured to, based on the cost function, calculate 0 as the cost for progress less than the threshold, and to calculate the cost as the power of the difference between the progress and the threshold, with a first value as the exponent, for progress equal to or greater than the threshold.

3. The behavior prediction device according to claim 2, wherein the prediction unit is configured to predict the time series of the progress based on a time limit, a target progress, a reward, the threshold value, the first value, and a second value indicating an amplification rate of the cost of the most recent behavior.

4. A program for causing a computer to function as the prediction unit provided in the behavior prediction device according to any one of claims 1 to 3.

Images