SYSTEM FOR SUPORTING USER'S BEHAVIOR

• US 20070106625A1
• Filed: 10/19/2006
• Published: 05/10/2007
• Est. Priority Date: 11/08/2005
• Status: Active Grant

• Current Assignee: International Business Machines Corporation
• Original Assignee: International Business Machines Corporation
• Inventors: Kashima, Hisashi
• Granted Patent: US 7,467,120 B2
• Time in Patent Office: Days
• US Class Current: 706/12
• CPC Class Codes: G06N 20/00 Machine learning
• CPC Class Codes: G06N 5/02 Knowledge representation; Symbolic representation

Abstract

Provided is a system 10 which supports a user'"'"'s behavior by generating a behavioral decision function indicating behavior to be adopted to a certain target. The system 10 includes: a data acquiring section 110 which acquires a cost caused as a result of adopting each of a plurality of behaviors to a target as training data for generating the behavioral decision function, the plurality of behaviors having already been adopted to the target; and a function generator 120 which generates, based on the training data, the behavioral decision function to minimize the expected shortfall of a cost to be obtained as a result of adopting the behavior to the target.

Claims

1. A system which supports a user'"'"'s behavior by generating a behavioral decision function indicating behavior to be adopted to a certain target, comprising:

a data acquiring section which acquires a cost caused as a result of adopting each of a plurality of behaviors to the target, as training data for generating the behavioral decision function, the plurality of behaviors having already been adopted to the target; and

a function generator which generates, based on the training data, the behavioral decision function to minimize expected shortfall of the cost to be obtained as a result of adopting the behavior to the target.

2. The system according to claim 1,

wherein the function generator includes:

a first calculator which calculates, based on the training data, the behavioral decision function to minimize an index value based on a total cost exceeding a value-at-risk, and stores the calculated function in a memory, in a case of using a given value as the value-at-risk of the cost;

a second calculator which reads the behavioral decision function calculated by the first calculator from the memory, and calculates, based on the training data, a value-at-risk of a cost caused as a result of adopting the behavior shown by the behavioral decision function, thus providing the result to the first calculator; and

a convergence judgment section which judges whether or not expected shortfall based on the index value has been converged to a value within a predetermined range, and

wherein the function generator outputs the behavioral decision function calculated by the first calculator on condition that the expected shortfall has been converged.

3. The system according to claim 2, wherein

the function generator further includes a third calculator which calculates the behavioral decision function to minimize an expected value of a cost based on the training data, and

the first calculator calculates the behavioral decision function to minimize the index value by subtracting the given value-at-risk from a cost corresponding to each behavior included in the training data and by providing the result for the third calculator.

4. The system according to claim 2, wherein

the first calculator calculates a parameter θ

of the behavioral decision function to minimize an index value C^E<sub>α

′</sub>

(θ

) which is calculated by the following equation 1 while setting the given value-at-risk to at α

′

,

the second calculator calculates α

^E_β

(θ

) calculated by the following equations 2 and 3 as a value-at-risk with respect to the calculated parameter θ

, and provides the result for the first calculator, and

the convergence judgment section judges whether or not a value calculated by the following equation 4 has converged to a value within a predetermined range,

equations 1 to 4 being defined as follows;

$\begin{array}{cc}\left(\mathrm{Equation}<br><br>\text{\hspace{1em}<br><br>}<br><br>1\right)<br><br><br><br>C_{{\mathrm{\alpha <br><br>}}^{\mathrm{\prime <br><br>}}}^E<br><br>\left(\mathrm{\theta <br><br>}\right)<br><br>\text{;<br><br>}=\frac{1}{N}<br><br>\underset{i=1}{\overset{N}{\sum <br><br>}}<br><br>\text{\hspace{1em}<br><br>}<br><br>{\left[c<br><br>\left(x^{\left(i\right)},h<br><br>\left(\mathrm{\theta <br><br>}\right)\right)-{\mathrm{\alpha <br><br>}}^{\mathrm{\prime <br><br>}}\right]}^{+}& \mathrm{Equation}<br><br>\text{\hspace{1em}<br><br>}<br><br>1\end{array}$

where targets of behaviors in the training data are denoted by x⁽¹⁾ to x⁽ⁿ⁾, the behavioral decision function is denoted by a function h for deciding behavior y to be adopted in accordance with a parameter θ

while setting a target x as input, the cost in the training data is expressed by c(x⁽ⁱ⁾, h(θ

)), and [x]⁺ denotes a function to return x when x is positive and 0 otherwise,

$\begin{array}{cc}\left(\mathrm{Equation}<br><br>\text{\hspace{1em}<br><br>}<br><br>2\right)<br><br><br><br>{\mathrm{\alpha <br><br>}}_{\mathrm{\beta <br><br>}}^E<br><br>\left(\mathrm{\theta <br><br>}\right)=c<br><br>\left(x^{\left(k^{\mathrm{\prime <br><br>}}\right)},h<br><br>\left(\mathrm{\theta <br><br>}\right)\right)& \mathrm{Equation}<br><br>\text{\hspace{1em}<br><br>}<br><br>2\\ \left(\mathrm{Equation}<br><br>\text{\hspace{1em}<br><br>}<br><br>\text{\hspace{1em}<br><br>}<br><br>3\right)<br><br><br><br>k^{\mathrm{\prime <br><br>}}=\min <br><br>\{k<br><br><br><br>\frac{1}{N}<br><br>\underset{i=1}{\overset{N}{\sum <br><br>}}<br><br>\text{\hspace{1em}<br><br>}<br><br>\left[I<br><br>\left(c<br><br>\left(x^{\left(i\right)},h<br><br>\left(\mathrm{\theta <br><br>}\right)\right)\ge <br><br>c<br><br>\left(x^{\left(k\right)},h<br><br>\left(\mathrm{\theta <br><br>}\right)\right)\right)\right]\le <br><br>1-\mathrm{\beta <br><br>}& \mathrm{Equation}<br><br>\text{\hspace{1em}<br><br>}<br><br>3\\ \left(\mathrm{Equation}<br><br>\text{\hspace{1em}<br><br>}<br><br>4\right)<br><br><br><br>F_{\mathrm{\beta <br><br>}}^E<br><br>\left(\mathrm{\theta <br><br>},{\mathrm{\alpha <br><br>}}^{\mathrm{\prime <br><br>}}\right)={\mathrm{\alpha <br><br>}}^{\mathrm{\prime <br><br>}}+\frac{1}{\left(1-\mathrm{\beta <br><br>}\right)<br><br>N}<br><br>\underset{i=1}{\overset{N}{\sum <br><br>}}<br><br>\text{\hspace{1em}<br><br>}<br><br>{\left[c<br><br>\left(x^{\left(i\right)},h<br><br>\left(\mathrm{\theta <br><br>}\right)\right)-{\mathrm{\alpha <br><br>}}^{\mathrm{\prime <br><br>}}\right]}^{+}& \mathrm{Equation}<br><br>\text{\hspace{1em}<br><br>}<br><br>4\end{array}$

where a function taking on 1 when a condition shown by an argument is satisfied and 0 when the condition is not satisfied is denoted by a function I, and that a probability of the generation of a cost exceeding a value-at-risk is denoted by a constant β

.

5. The system according to claim 3, wherein the first calculator calculates a cost c′

⁽ⁱ⁾ by subtracting α

′

being the given value-at-risk from a cost c⁽ⁱ⁾, with respect to a target x⁽ⁱ⁾ and behavior y in the training data, by use of the following equation 5, and provides the result for the third calculator,

the third calculator calculates the behavioral decision function to minimize a cost C′

^E<sub>α

′</sub>

(θ

) expressed by the following equation 6, equations 5 and 6 being defined as follows;

$\begin{array}{cc}\left(\mathrm{Equation}<br><br>\text{\hspace{1em}<br><br>}<br><br>5\right)<br><br><br><br>c^{\mathrm{\prime <br><br>}<br><br>\left(i\right)}<br><br>\left(x^{\left(i\right)},y\right)={\left[c^{\left(i\right)}<br><br>\left(x^{\left(i\right)},y\right)-{\mathrm{\alpha <br><br>}}^{\mathrm{\prime <br><br>}}\right]}^{+}& \mathrm{Equation}<br><br>\text{\hspace{1em}<br><br>}<br><br>5\\ \left(\mathrm{Equation}<br><br>\text{\hspace{1em}<br><br>}<br><br>\text{\hspace{1em}<br><br>}<br><br>6\right)<br><br><br><br>C_{{\mathrm{\alpha <br><br>}}^{\mathrm{\prime <br><br>}}}^{\mathrm{\prime <br><br>}<br><br>E}<br><br>\left(\mathrm{\theta <br><br>}\right)=\frac{1}{N}<br><br>\underset{i=1}{\overset{N}{\sum <br><br>}}<br><br>\text{\hspace{1em}<br><br>}<br><br>c^{\mathrm{\prime <br><br>}<br><br>\left(i\right)}<br><br>\left(x^{\left(i\right)},y\right)& \mathrm{Equation}<br><br>\text{\hspace{1em}<br><br>}<br><br>6\end{array}$

6. The system according to claim 1, wherein the system is a system which supports a decision on a treatment policy for a patient having a certain test result, wherein

the behavioral decision function is a function which decides a treatment policy for a patient,

the data acquiring section acquires the scale of a loss, which is caused on each of a plurality of patients, as the training data, the plurality of patients having already been treated, and

the function generator generates, based on the training data, the behavioral decision function to minimize expected shortfall of a loss caused by a treatment.

7. A method which supports a user'"'"'s behavior by generating a behavioral decision function showing behavior to be adopted to a target, comprising the steps of:

acquiring a cost caused as a result of adopting each of a plurality of behaviors to the target, as training data for generating the behavioral decision function, the plurality of behaviors having already been adopted to the target; and

generating, based on the training data, the behavioral decision function to minimize expected shortfall of a cost to be obtained as a result of adopting the behavior to the target.

8. A program causing an information processor to function as a system supporting a user'"'"'s behavior by generating a behavioral decision function showing behavior to be adopted to a target, the program causing the information processor to function as:

a data acquiring section which acquires, as training data for generating the behavioral decision function, a cost caused as a result of adopting each of a plurality of behaviors to the target, the plurality of behaviors having already been adopted to the target; and

a function generator which generates, based on the training data, the behavioral decision function to minimize expected shortfall of a cost to be obtained as a result of adopting the behavior to the target.