IP Library Granted Patent US 12711360
Granted Patent B2
US 12711360 · App. 18/287,132 · Granted Aug 18, 2026

Learning device, learning method, and learning program

Inventors: Dai Kubota (Tokyo, JP); Riki Eto (Tokyo, JP)
Assignee: NEC CORPORATION
G06N3/048G06N3/092
View Patent ↗
Loading inventors, assignments & file history…
Monitor This Case
Get email alerts when status or documents change.
Order Certified Copies
Most orders are placed with the USPTO same day — all within 24 business hours.
Order via The Patent Place →
Pre-filled with this patent's details
Quick Facts
Patent No.
US 12711360
App. No.
18/287,132
Granted
Aug 18, 2026
Kind
B2
Abstract

The input means 81 accepts input of trajectory data indicating the subject's decision-making history. The learning means 82 performs inverse reinforcement learning using the trajectory data. The output means 83 outputs a reward function and a constraint condition derived by inverse reinforcement learning. Here, the learning means 82 performs inverse reinforcement learning based on distribution of the trajectory data calculated using a differentiable function that indicates distribution of the constraint condition.

Claims (41)

1 . A learning device comprising:

a memory storing instructions; and

one or more processors configured to execute the instructions to:

accept input of trajectory data indicating the subject's decision-making history;

perform inverse reinforcement learning using the trajectory data; and

output a reward function and a constraint condition derived by inverse reinforcement learning,

wherein the inverse reinforcement learning is performed based on distribution of the trajectory data calculated using a differentiable function that indicates distribution of the constraint condition.

2 . The learning device according to claim 1 , wherein the processor is configured to execute the instructions to:

accept input of a probabilistic model that assumes a distribution of trajectory data derived by maximum entropy principle based on the reward function and the constraint condition;

calculate a gradient of a log-likelihood based on a probability model and trajectory data, and update the reward function and the constraint condition to maximize the log-likelihood;

the probability model is defined as product of a first function indicating distribution based on the reward function expressed using a first parameter and a second function indicating distribution based on the constraint condition expressed using a second parameter; and

update the first parameter and the second parameter to maximize the log-likelihood.

3 . The learning device according to claim 2 , wherein the processor is configured to execute the instructions to:

determine whether the first parameter and the second parameter have converged; and

when the parameters are not determined to have converged, calculate the gradient of the log-likelihood based on the probability model defined by the updated first parameter and second parameter and the trajectory data, and update the first parameter and second parameter to maximize the log-likelihood.

4 . The learning device according to claim 1 , wherein

the second function is defined as a continuous function differentiable by the second parameter such that the more each constraint condition is satisfied, the closer it approaches 1, and the less each such constraint condition is satisfied, the closer it approaches 0.

5 . The learning device according to claim 2 , wherein

the probability model is defined as a function that takes a larger probability value the larger a value of the reward function and a larger probability value the more constraint condition is satisfied.

6 . The learning device according to claim 1 , wherein

the second function is defined by the total power of a sigmoid function whose argument is a constraint condition that indicates a non-negative value when each constraint condition is satisfied.

7 . A learning method comprising:

accepting input of trajectory data indicating the subject's decision-making history;

performing inverse reinforcement learning using the trajectory data;

outputting a reward function and a constraint condition derived by inverse reinforcement learning; and

inverse reinforcement learning is performed based on distribution of the trajectory data calculated using a differentiable function that indicates distribution of the constraint condition.

8 . The learning method according to claim 7 , further comprising:

accepting input of a probabilistic model that assumes a distribution of trajectory data derived by maximum entropy principle based on the reward function and the constraint condition;

a probability model is defined as product of a first function indicating distribution based on the reward function expressed using a first parameter and a second function indicating distribution based on the constraint condition expressed using a second parameter;

calculating a gradient of a log-likelihood based on the probability model and trajectory data, and updating the reward function and the constraint condition to maximize the log-likelihood; and

updating the first parameter and the second parameter to maximize the log-likelihood.

9 . A non-transitory computer readable information recording medium storing a learning program, when executed by a processor, that performs a method for:

accepting input of trajectory data indicating the subject's decision-making history;

performing inverse reinforcement learning using the action trajectory data; and

outputting a reward function and a constraint condition derived by inverse reinforcement learning,

wherein inverse reinforcement learning is performed based on distribution of the trajectory data calculated using a differentiable function that indicates distribution of the constraint condition, in the learning processing.

10 . The non-transitory computer readable information recording medium according to claim 9 , wherein

input of a probabilistic model that assumes a distribution of trajectory data derived by maximum entropy principle based on the reward function and the constraint condition is accepted,

a gradient of a log-likelihood based on a probability model and trajectory data is calculated, and the reward function and the constraint condition are updated to maximize the log-likelihood,

the probability model is defined as product of a first function indicating distribution based on the reward function expressed using a first parameter and a second function indicating distribution based on the constraint condition expressed using a second parameter, and

the first parameter and the second parameter are updated to maximize the log-likelihood.