INFORMATION PROCESSING APPARATUS, INFORMATION PROCESSING METHOD AND PROGRAM

• US 20130066817A1
• Filed: 08/20/2012
• Published: 03/14/2013
• Est. Priority Date: 09/09/2011
• Status: Active Grant

• Current Assignee: Sony Corporation (Sony Group Corp.)
• Original Assignee: Sony Corporation (Sony Group Corp.)
• Inventors: KAWAMOTO, Kenta
• Granted Patent: US 8,924,317 B2
• Time in Patent Office: Days
• US Class Current: 706/12
• CPC Class Codes: G06N 20/00 Machine learning
• CPC Class Codes: G06N 7/01 Probabilistic graphical models, e.g. probabilistic networks

Abstract

There is provided an information processing apparatus including an information amount gain calculation unit that, on the assumption that a factor that could cause a state transition in a state transition model occurs and the state transition takes place as a result of an occurrence of the factor, determines an information amount gain, which is an information amount obtained by causing the factor to occur regarding a causal relationship between the factor and the state transition and observing a result thereof, an evaluation value calculation unit that determines an evaluation value to evaluate effectiveness of causing each of the factors to occur based on the information amount gain to acquire the causal relationship between the factor and the state transition, and a decision unit that decides the factor to be occurred based on the evaluation value.

Claims

1. An information processing apparatus, comprising:

an information amount gain calculation unit that, on the assumption that a factor that could cause a state transition in a state transition model occurs and the state transition takes place as a result of an occurrence of the factor, determines an information amount gain, which is an information amount obtained by causing the factor to occur regarding a causal relationship between the factor and the state transition and observing a result thereof;

an evaluation value calculation unit that determines an evaluation value to evaluate effectiveness of causing each of the factors to occur based on the information amount gain to acquire the causal relationship between the factor and the state transition; and

a decision unit that decides the factor to be occurred based on the evaluation value.

2. The information processing apparatus according to claim 1,

wherein the information amount gain calculation unit

determines an inter-distribution distance indicating a divergence between a probability distribution of a base probability as a probability with which the state transition takes place in a predetermined state and a probability distribution of a factor conditional probability as a probability with which a predetermined factor occurs and the state transition takes place in the predetermined state as a causal relationship amount indicating a degree of the causal relationship between the predetermined factor and the state transition in the predetermined state, and

causes the predetermined factor to occur to determine a difference between the causal relationship amount after observing a result thereof and the causal relationship amount before the predetermined factor occurs as the information amount gain.

3. The information processing apparatus according to claim 2,

wherein the inter-distribution distance is a KL (Kullback-Leibler) divergence of the probability distribution of the base probability and the probability distribution of the factor conditional probability.

4. The information processing apparatus according to claim 3,

wherein the probability with which the state transition takes place follows a beta distribution.

5. The information processing apparatus according to claim 4,

wherein an expected value of the base probability as the probability with which the state transition takes place in a predetermined state to another predetermined state is represented as B_T/B_S by using the number of times B_S of being in the predetermined state and the number of times B_T that the state transition takes place from the predetermined state to the other predetermined state,

wherein the expected value of the factor conditional probability as the probability with which the predetermined factor occurs and the state transition in the predetermined state to the other predetermined state is represented as N_T/N_S by using the number of times N_S of the occurrence of the predetermined factor in the predetermined state, the number of times of having been in the predetermined state being B_S, and the number of times N_T that the state transition takes place from the predetermined state to the other predetermined state when the predetermined factor occurs,

wherein the base probability follows the beta distribution of a beta function B(B_T+1,B_S−

B_T+1), and

wherein the factor conditional probability follows the beta distribution of a beta function B(N_T+1, N_S−

N_T+1).

6. The information processing apparatus according to claim 5,

wherein the evaluation value calculation unit determines, as the evaluation value to evaluate the effectiveness of causing the predetermined factor to occur, a weighted added value obtained by adding the expected value of the information amount gain by assigning a transition probability of the state transition of the state transition model as a weight to acquire the causal relationship between the predetermined factor and the state transition from the predetermined state.

7. The information processing apparatus according to claim 6,

wherein the decision unit determines a factor path as a series of the factors until each of the factors occurs from a current factor as the factor currently occurring and decides the factor path with a maximum average value of an added value of the evaluation value of the factor as a causal acquisition plan as a series of the factors caused to acquire the causal relationship.

8. The information processing apparatus according to claim 7,

wherein the decision unit determines the factor path by limiting the number of factors constituting the factor path.

9. The information processing apparatus according to claim 7,

wherein the decision unit determines a series of factors containing no loop as the factor path.

10. The information processing apparatus according to claim 7,

wherein the decision unit determines a shortest path from the current factor to each of the factors as the factor path.

11. The information processing apparatus according to claim 7,

wherein the state transition model is an HMM (Hidden Markov Model), and

wherein, when a current state of a first HMM as one HMM of a plurality of HMMs that have learned by using observed values observed by an agent and the current state of a second HMM of at least one other HMM are recognized by using the observed values,

the information amount gain calculation unit assumes that being in each state of the second HMM is each of the factors and the state transition from the current state of the first HMM occurs as a result of being in each state of the second HMM to determine the information amount gain obtained by being in each state of the second HMM regarding the causal relationship between each state of the second HMM and the state transition of the first HMM,

the evaluation value calculation unit determines, as the evaluation value to evaluate the effectiveness of being in each state of the second HMM, the weighted added value obtained by adding the expected value of the information amount gain by assigning the transition probability of the state transition of the first HMM as the weight to acquire the causal relationship between the states of the second HMM and the state transition of the first HMM, and

the decision unit determines a series of states to reach each state of the second HMM from the current state of the second HMM as the factor path and decides the series of states of the second HMM as the factor path with the maximum average value of the added value of the evaluation value as the causal acquisition plan.

12. The information processing apparatus according to claim 11,

wherein, when the state transition of the second HMM is not restricted, the series to reach each of the states of the second HMM from the current state of the second HMM is determined as the factor path.

13. The information processing apparatus according to claim 11,

wherein the second HMM is an HMM that has learned movement of an arm of the agent capable of taking action to move the arm,

wherein the first HMM is an HMM that has learned movement of a moving object that is moved by being touched by the arm, and

wherein the agent moves the arm in such a way that the states of the second HMM constituting the series as the causal acquisition plan sequentially become the current state.

14. An information processing method, comprising:

on the assumption that a factor that could cause a state transition in a state transition model occurs and the state transition takes place as a result of an occurrence of the factor, determining an information amount gain, which is an information amount obtained by causing the factor to occur regarding a causal relationship between the factor and the state transition and observing a result thereof;

determining an evaluation value to evaluate effectiveness of causing each of the factors to occur based on the information amount gain to acquire the causal relationship between the factor and the state transition; and

deciding the factor to be occurred based on the evaluation value.

15. A program for causing a computer to function as:

an information amount gain calculation unit that, on the assumption that a factor that could cause a state transition in a state transition model takes place and the state transition occurs as a result of an occurrence of the factor, determines an information amount gain, which is an information amount obtained by causing the factor to occur regarding a causal relationship between the factor and the state transition and observing a result thereof;

an evaluation value calculation unit that determines an evaluation value to evaluate effectiveness of causing each of the factors to occur based on the information amount gain to acquire the causal relationship between the factor and the state transition; and

a decision unit that decides the factor to be occurred based on the evaluation value.