Reinforcement learning system专利检索

BACKGROUND OF THE INVENTION

1. Field of the Invention

The present invention relates to a reinforcement learning system for making an agent learn an action policy for executing a task.

2. Description of the Related Art

In an attempt to reduce a learning time for obtaining a behavior suitable for a task, there has been disclosed an art (refer to Japanese Patent Laid-open No. 2005-078516) which determines action policies for an agent from learning results obtained by a plurality of learning devices based on the surrounding state and selects an action policy from the plural action policies based on the learning performance of each learning device.

According to the conventional art, only the learning result from a single learning device is utilized, thereby, it is difficult for the agent to execute such a task that involves complicated actions which may be achieved by making full use of the learning results from the other learning devices.

SUMMARY OF THE INVENTION

The present invention has been accomplished in view of the aforementioned problems, and it is therefore an object of the present invention to provide a reinforcement learning system capable of making an agent learn an action policy for executing a task involving complicated actions.

To attain an object described above, the reinforcement learning system for making an agent learn an action policy for executing a task according to the present invention comprises: an environment recognizing device configured to recognize a first to an nth (n≧2) state variables representing an environment; n numbers of jth (j=1, 2, . . . n) learning devices configured to calculate a jth reward based on a jth state variable recognized by the environment recognizing device, calculate a jth value according to a jth value function based on the jth state variable, calculate a jth error based on the jth reward and the jth value, and properly modify respectively the jth value function based on the jth error; and an action policy determining device configured to determine an action policy the agent should take based on a part or the entire part of the jth reward calculated by the n numbers of the jth learning device, wherein an (i+1)th (i=1, 2, . . . n−1) learning device is configured to calculate an (i+1)th reward based on a value of an ith reward function related to an ith state variable and a value of an ith value gradient function which is a temporal differential of an ith value function.

According to the reinforcement learning system of the present invention, the value of the ith value gradient function can be utilized in the learning performed by the (i+1)th (i=1, 2, . . . , n−1) learning device. Herein, the ith value gradient function is a temporal differential of the ith value function which is defined according to the ith reward obtained from the environment and is also a learning result obtained by the ith learning device. Note that the ith state variable is a conception including scalar and vector.

According thereto, in the learning performed by the nth learning device, the learning result performed by the (n−1)th learning device can be utilized directly and the learning results by the learning devices prior to the (n−2)th learning devices can be utilized indirectly. In other words, all the learning policies represented by the first to the nth reward functions for the first to the nth learning devices, respectively, are expressed by the nth value function which is the learning result by the nth learning device. Consequently, an action policy which should be taken by the agent can be determined based on a part or the entire part of the first to the nth rewards.

Therefore, an action policy defined according to the variation behaviors of a plurality of state variables can be learned by the agent to execute a task involving complicated actions.

It is acceptable to configure the (i+1)th learning device to evaluate the (i+1)th reward lower as the ith value gradient function goes greater toward a negative side.

According to the reinforcement learning system with the mentioned configuration, in a case where the values of the ith value function decrease in a time series, in other words, when the environment defined according to the action policy of the agent is contradicted to the learning policy of the ith learning device, the (i+1)th reward calculated by the (i+1)th learning device is evaluated lower.

Accordingly, an action policy contradicted to the learning policy of the ith learning device can be prevented from being adopted as the action policy of the agent even though the environment is in compliance with the learning policy of the (i+1)th learning device. Specifically, even though an action policy is proper in view of the variation behaviors of a part of the state variables, if it is apparently improper in view of the variation behaviors of the other left state variables, it will not be adopted by the agent as an action policy thereof.

Therefore, an action policy defined according to the variation behaviors of the plural state variables can be learned by the agent to execute a task involving complicated actions.

It is acceptable to configure the (i+1)th learning device as to evaluate the (i+1)th reward irrelevant to the value of the ith value gradient function only on an exceptional situation where the ith value gradient function has a positive value.

According to the reinforcement learning system with the mentioned configuration, in a case where the values of the ith value function increase in a time series, in other words, when the environment defined according to the action policy of the agent is in compliance with the learning policy of the ith learning device, the (i+1)th reward is evaluated by the (i+1)th learning device irrelevant to the value of the ith value gradient function as an exception.

Accordingly, the situation where the (i+1)th reward is evaluated excessively high due to the reason that the environment is in compliance with the learning policy of the ith learning device and an action policy contradicted to the learning policy of the (i+1)th learning device is adopted as the action policy for the agent unexpectedly can be avoided. Specifically, such situation can be avoided that the action policy contradicted to the learning policy of a subsequent learning device is adopted due to the reason that the learning result by a previous learning device is excessively emphasized in the learning by the subsequent learning device.

Therefore, an action policy defined according to the variation behaviors of the plural state variables can be learned by the agent to execute a task involving complicated actions.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 is a diagram illustrating a configuration of a reinforcement learning system of the present invention.

FIG. 2 is a diagram illustrating a configuration of a robot served as an agent.

FIG. 3 is a flow chart illustrating an action policy determining method performed by the reinforcement learning system.

FIG. 4 is an explanatory diagram illustrating a task executed by the agent.

FIG. 5 is an explanatory diagram illustrating different reward functions.

FIG. 6 is an explanatory diagram illustrating a value function and a value gradient function which is a time differential thereof.

FIG. 7 is an explanatory diagram illustrating a task executed according to a determined action policy.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

Hereinafter, an embodiment of a reinforcement learning system according to the present invention will be described in detail with reference to the drawings.

The configuration of the reinforcement learning system will be firstly described. The reinforcement learning system 1 illustrated in FIG. 1 is composed of an electronic control unit (constituted from a CPU, a ROM, a RAM, an I/O circuit and the like) or a computer. The electronic control unit or the computer is mounted in a robot R illustrated in FIG. 2. The robot R is served as an agent in the present invention.

The CPU of the computer constituting the reinforcement learning system 1 retrieves a reinforcement learning program (software) from memory and performs various processing such as a learning processing (to be described hereinafter) according to the retrieved reinforcement learning program. The reinforcement learning program may be installed in the computer by means of a software recording medium such as CD or DVD, or may be downloaded to the computer through a network or a satellite from a server in a condition when the robot R sends a request signal to the server.

The reinforcement learning system 1 includes an environment recognizing device 11, a first learning device 121, a second learning device 122, and an action policy determining device 13.

The environment recognizing device 11 is configured to recognize a first state variable and a second state variable representing an environment based on output signals from an internal state sensor set S1 and an external state sensor set S2, both of which are mounted in the robot R.

The first learning device 121 is configured to calculate a first reward r1(t) according to a first reward function r1 based on the first state variable (details of which will be described hereinafter) recognized by the environment recognizing device 11 and a first value V1(t) according to a first value function V1 based on the first state variable, and thereafter to calculate a first TD error (first error) δ1(t) corresponding to the first reward r1(t) and the first value V1(t). The first learning device 121 is further configured to modify properly a first value function V1 based on the first error δ1(t). The first reward function r1 is defined according to a learning policy which makes a position p of a hand B4 achieve a desired position p0 while keeping down a kinetic energy E thereof.

The second learning device 122 is configured to calculate a second reward r2(t) based on a value of a second reward function r2 which is related to the second state variable (details of which will be described hereinafter) recognized by the environment recognizing device 11 and a value of a first value gradient function dV1/dt which is a temporal differential of the first value function V1 and a second value V2(t) according to a second value function V2 based on the second state variable, and thereafter to calculate a second TD error (second error) δ2(t) based on the second reward r2(t) and the second value V2(t). The second learning device 122 is further configured to modify properly a second value function V2 based on the second error δ2(t). The second reward function r2 is defined according to a learning policy which makes a posture q of a hand B4 achieve a desired posture q0.

The action policy determining device 13 is configured to determine a first output (details of which will be described hereinafter) obtained according to a first output function as a part of an action policy that should be taken by the robot (agent) R based on an environment recognition result from the environment recognizing device 11. Similar to the first value function V1, the first output function is properly modified based on the first error δ1(t).

The action policy determining device 13 is configured to determine a second output (details of which will be described hereinafter) obtained according to a second output function as a part of the action policy that should be taken by the robot (agent) R based on the environment recognition result from the environment recognizing device 11. Similar to the second value function V2, the second output function is properly modified based on the second error δ2(t). The sum of the first output and the second output is applied to the robot R.

The reinforcement learning system 1 is configured to control actions of the robot R through controlling operations of an actuator 100 according to the action policy determined by the action policy determining device 13.

It should be noted that “a component of the present invention “recognizes” information” means that the component performs a possible information processing on a piece of information to prepare the piece of information ready for other information processing, for example, the component searches the piece of information in a database; retrieves the piece of information from a storing device such as a memory; measures, calculates, estimates, determines the piece of information based on output signals from a sensor or the like; and stores in memory or the like the determined information or the like.

The robot R is a humanoid robot. Similar to a human being, the robot R is provided with a main body B0, a head B1 disposed at an upper end of the main body B0, a pair of arms B2 extended from an upper portion of the main body B0 to left and right sides, respectively, a hand B4 disposed at an end portion of each arm B2, a pair of left and right legs B3 extended downward from a lower end of the main body B0, and a foot B5 disposed at an end portion of each leg B3.

As disclosed by Japanese Patent Laid-open No. 03/090978 and No. 03/090979 Republished, according to a force transmitted from the actuator, the robot R can stretch and/or bend the arms B2 and legs B3 through a plurality of joint mechanisms corresponding to a plurality of joints of a human being, such as the shoulder joint, the elbow joint, the wrist joint, the hip joint, the knee joint, the ankle joint and the like.

The arm B2 includes a first arm link and a second arm link. The first arm link is connected to the main body B0 via the shoulder joint mechanism. One end of the second arm link is connected to the end portion of the first arm link via the elbow joint mechanism, and the other end thereof is connected to the end portion of the hand B4 via the wrist joint mechanism.

The leg B3 includes a first leg link and a second leg link. The first leg link is connected to the main body B0 via the hip joint mechanism. One end of the second leg link is connected to the end portion of the first leg link via the knee joint mechanism, and the other end thereof is connected to the foot B5 via the ankle joint mechanism.

The robot R can move around autonomously according to iterated movements involving leaving off the floor and stepping on the floor of the left and right legs B3, respectively.

As aforementioned, the robot R is mounted with the internal state sensor set S1 and the external state sensor set S2.

The internal state sensor set S1 includes a GPS determining device or an acceleration sensor for determining a position (position of the center of gravity) of the robot R, a gyro sensor for determining a posture of the main body B0, rotary encoders for determining bending angles and the like of each joint mechanism.

The external state sensor set S2 includes a stereo image sensor, an active distance sensor and the like. The stereo image sensor is mounted in the head B1 and is configured to determine an object, such as a cup, to be held by the robot R. The active distance sensor is mounted in the main body B0 and uses infrared lights to measure the distance actively.

The functions of the reinforcement learning system 1 with the mentioned configurations will be described hereinafter.

The reinforcement learning system 1 is used to learn the optimum action policy for making the robot R execute a task, for example, to hold an object such as a cup with the hand B4 by moving the arm B2 of a single side. For simple explanation, as illustrated in FIG. 4, the action policy is presumed to be learned at a constraint condition that the movements of the arm B2 and the hand B4 are constrained in a X-Y plane with the position of the shoulder joint as the origin O. The action policy that should be taken by the robot R may be determined according to simulations or experiments with the robot (real product) R used therein.

First, an environment is recognized by the environment recognizing device 11 (FIG. 3/STEP 002).

Specifically, a two-dimensional position p(t)=(x(t), y(t)) of the hand B4 is recognized as the first state variable, and a posture q(t) of the hand B4 is recognized as the second state variable.

The position p(t) of the hand B4 is determined according to geometrical relationships based on a length L1 of the first arm link, a length L2 of the second arm link, a length L3 of the hand B4, the shoulder joint angle q1(t), the elbow joint angle q2(t) and the wrist joint angle q3(t) (refer to FIG. 4). The data of the lengths L1, L2 and L3 are preliminarily stored in memory.

The shoulder joint angle q1(t) refers to an angle formed by the first arm link with respect to the frontal plane (x axis) viewed above from the robot R. The elbow joint angle q2(t) refers to a bending angle of the second arm link with respect to the first arm link viewed above from the robot R. The wrist joint angle q3(t) refers to a bending angle of the hand B4 with respect to the second arm link viewed above from the robot R.

The posture q(t) of the hand B4 is determined by adding up the shoulder joint angle q1(t), the elbow joint angle q2(t) and the wrist joint angle q3(t) (refer to FIG. 4). The joint angles q1(t), q2(t) and q3(t) can be determined according to output signals from rotary encoders (constituting the internal state sensor set S1) disposed at the shoulder joint, the elbow joint and the wrist joint of the robot, respectively.

The position p(t) of the hand (the end portion) B4 may be determined according to the stereo image sensor (constituting the external state sensor set S2) mounted in the head B1. The consumed electricity E(t) for transiting the hand B4 may be determined according to a data table or a computation equation representing a relationship between the rotation amount and the consumed electricity for each joint actuator based on variation amounts of the joint angles q1(t), q2(t) and q3(t).

The first state variable p(t) and the second state variable q(t) are input to the first learning device 121 and the second learning device 122, respectively. If common information is input, it is possible to compare the information between the first learning device 121 and the second learning device 122. Moreover, the second state variable q(t) is input to the action policy determining device 13.

Subsequently, the first reward r1(t) is evaluated by the first learning device 121 based on the first state variable p(t) according to the first reward function r1 expressed by the equation (1) (FIG. 3/STEP 102).

r1(t)=exp(−αp|p(t)−p0|2)−βE(t)  (1)

Herein, p0 denotes the desired position of the hand B4, αp and β are positive coefficients. As illustrated in FIG. 5, the first reward function r1(t) holds a mountain-like shape which has the desired position p0 as the apex and levels down as the position p of the hand B4 moves away from the desired position p0.

As obviously seen from the first term at the right side of the equation (1), the closer the present position p(t) of the hand B4 is to the desired position p0, the higher the first reward r1(t) is evaluated. As obviously seen from the second term at the right side of the equation (1), the smaller the kinetic energy E(t) is, the higher the first reward r1(t) is evaluated. The kinetic energy E(t) can be calculated according to a data table or a computation equation representing a relationship between the rotation amount and the kinetic energy of each joint actuator based on variation amounts of the joint angles q1(t), q2(t) and q3(t).

The first value V1(t) is calculated by the first learning device 121 according to the first value function V1 based on the first state variable p(t) (FIG. 3/STEP 104).

The first value function V1 is determined according to NGNet which actually complies with the thinking of adding a weight to the frequency of occurrence of data based on the first reward function r1 (for example, refer to Reinforcement Learning in Continuous Time and Space, Kenji Doya, Neural Computation, 12(1), 219-245(2000)). It is acceptable to determine the first value function V1 according to a publicly known method different from NGNet, such as RBF, neural network, CMAC or the like.

Then, based on the first reward r1(t) and the first value V1(t), the first TD error δ1(t) is calculated according to the equation (2) (FIG. 3/STEP 106).

δ1(t)=r1(t)+γ1V1(t+1)−V1(t)  (2)

Herein, γ1 represents a discount factor in the Bellman equation adopted by the first learning device 121. The appropriateness of an action policy taken by the robot R previously is evaluated according to the magnitude of the first TD error δ1(t). As aforementioned, the first value function V1 and the first output function are properly modified based on the first TD error δ1(t). Thus, there is a tendency that the modification amounts thereof become smaller when the action policy taken by the robot R previously is the same as the expected, and on the other hand, the modification amounts thereof become greater when the action policy taken by the robot R previously is different from the expected.

Thereafter, based on the posture (the second state variable) q(t) of the hand B4 which comes as the recognition result of the environment recognizing device 11, the first output is determined as the action policy of the robot R by the action policy determining device 13 according to the first output function (FIG. 3/STEP 108). Specifically, the torque T1 around the shoulder joint and the torque T2 around the elbow joint of the robot R, which make the integrated value of the first reward r1(t) become maximum, are determined as the first output.

Subsequently, based on the value of the second reward function in accordance with the second state variable q(t) and the value of the first value gradient function (temporal differential of the first value function V1) dV1/dt, the second reward r2(t) is evaluated according to the equation (3) by the second learning device 122 (FIG. 3/STEP 202).

r2(t)=exp(−αq|q(t)−q0|2)+ηn·Y(−x),

x=dV1/dt,

Y(x)≡(x≧0) or 0(x