Robotic learning and evolution apparatus

• US 8,793,205 B1
• Filed: 09/20/2012
• Issued: 07/29/2014
• Est. Priority Date: 09/20/2012
• Status: Active Grant

• Current Assignee: Brain Corporation
• Original Assignee: Brain Corporation
• Inventors: Fisher, Dimitry, Izhikevich, Eugene
• Primary Examiner(s): Starks, Wilbert L
• Application Number:
  US13/623,826
• Time in Patent Office: 677 Days
• Field of Search: None
• US Class Current: 706/25
• CPC Class Codes: G05D 1/0221 involving a learning process
• CPC Class Codes: G06N 20/00 Machine learning
• CPC Class Codes: G06N 3/008 based on physical entities controlled by simulated intelligence so as to replicate intelligent life forms, e.g. based on robots replicating pets or humans in their appearance or behaviour

Abstract

Apparatus and methods for implementing robotic learning and evolution. An ecosystem of robots may comprise robotic devices of one or more types utilizing artificial neuron networks for implementing learning of new traits. A number of robots of one or more species may be contained in an enclosed environment. The robots may interact with objects within the environment and with one another, while being observed by the human audience. In one or more implementations, the robots may be configured to ‘reproduce’ via duplication, copy, merge, and/or modification of robotic. The replication process may employ mutations. Probability of reproduction of the individual robots may be determined based on the robot'"'"'s success in whatever function trait or behavior is desired. User-driven evolution of robotic species may enable development of a wide variety of new and/or improved functionality and provide entertainment and educational value for users.

Claims

1. A robotic apparatus, comprising:

one or more processors configured to execute computer program modules, the computer program modules being executable to effectuate a neural network, the neural network being characterized by a configuration vector, the configuration vector comprising a combination of at least a portion of a first parent vector and at least a portion of second parent vector;

wherein;

the first parent vector is based on an operation of a first neural network of a first parent in accordance with a learning process, the operation of the first neural network being characterized by a first network configuration vector configured based on the first parent achieving a first task;

the second parent vector is based on an operation of a second neural network of a second parent in accordance with a learning process, the operation of the second network being characterized by a second network configuration vector configured based on the second parent achieving a second task; and

the combination is based on a selection received from at least one agent responsive to the first parent achieving the first task and the second parent achieving the second task, the selection being indicative of the first neural network and the second neural network.

2. The apparatus of claim 1, wherein the configuration vector comprises a plurality of weights associated with synaptic connections of the neural network.

3. The apparatus of claim 1, wherein individual weights associated of the configuration vector are based on a linear combination of individual weights of the first parent vector and individual weights of the second parent vector.

4. The apparatus of claim 3, wherein the combination comprises a random operation configured to introduce a random component into individual weights of the configuration vector.

5. A computer-implemented method of operating a robotic apparatus comprising network of a plurality of spiking neurons, the method being performed by one or more processors configured to execute computer program modules, the method comprising:

operating the network in accordance with a learning process configured to cause the apparatus to perform a task, the operating being characterized by a network state information; and

based on a completion of the task, transferring at least a portion of the network state information to another robotic apparatus;

wherein the transferring is configured to enable the other robotic apparatus to perform the task.

6. The method of claim 5, wherein:

the network comprises a plurality of connections configured to communicate spikes between the plurality of neurons, individual ones of the plurality of connections being characterized by connection weight; and

the operating network in accordance with the learning process comprises adjusting one or more of the connection weights in accordance with a connection plasticity mechanism.

7. The method of claim 6, wherein:

the completion of the task by the robotic apparatus is characterized by an outcome having a performance value associated therewith, the outcome being associated with a desired network state;

the adjusting the one or more of the connection weights is configured to transition present network state to the desired network state; and

the present network state having a present performance value associated therewith, the present performance value being lower than the performance value.

8. The method of claim 7, wherein:

the other robotic apparatus comprises another spiking network configured to operate in accordance with the learning process, characterized by another present performance value; and

the enabling of the other robotic apparatus to perform the task is characterized by the other network attaining another target performance that is greater than the other present performance.

9. The method of claim 8, wherein:

the performance value is within an range of performance values;

the other robotic apparatus comprises another spiking network configured to operate in accordance with the learning process; and

the enabling of the other robotic apparatus to perform the task is characterized by the other network attaining another target performance that is within the established range.

10. The method of claim 9, wherein:

the portion of the network state information comprises a vector comprising a plurality of weights of individual ones of the plurality of connections;

the other network comprises another plurality of connections characterized by another vector of connection weights; and

the transferring is configured to replicate the at least a portion of the vector within the other vector without relying on individual adjustment of weights of the other vector by the learning process.

11. The method of claim 10, wherein the transferring is configured to copy the vector into the other vector.

12. The method of claim 10, wherein:

the transferring is configured to apply a mathematical transformation to individual elements of the vector to obtain the transformed vector; and

the other vector comprises the transformed vector.

13. The method of claim 10, wherein the transformation comprises a random operation configured to introduce a random component to individual weight elements of the vector.

14. A computer-implemented method of updating a state of a neural network device, the method being performed by one or more processors configured to execute computer program modules, the method comprising:

establishing a data connection to a cloud server apparatus;

facilitating browsing of a plurality of state files via a user interface, the state files being stored on the cloud server apparatus;

receiving a selection of an individual one of the plurality of state files;

establishing a second data connection to the neural network device; and

causing the extraction and application of a state by the neural network device, the state being described at least in part by the individual one of the plurality of state files.

15. A computer-implemented method of updating a state of a neural network device, the method being performed by one or more processors configured to execute computer program modules, the method comprising:

observing operation of two or more neural network devices, the operation having a selection threshold associated therewith;

responsive to performance of another neural network device of the two more neural network devices being above the threshold, reproducing a configuration of the other neural network device by transferring at least a portion of neural network image of the other neural network device into the neural network device; and

responsive to performance of another neural network device of the two more neural network devices being below the threshold, preventing reproduction of the configuration of the other neural network device.

16. The method of claim 15, wherein:

the selection threshold is associated with completion of a task by the other neural network device;

the performance being above the threshold is based on a completion of the task; and

the performance being below the threshold is based on an unsuccessful attempt to complete the task.

17. The method of claim 15, wherein

individual ones of the two or more neural network devices comprise autonomous racing robotic apparatus;

the task comprises a race between the autonomous racing robotic apparatus absent user intervention; and

the selection threshold comprises winning the race.

18. The method of claim 15, wherein

the configuration of the other neural network comprises a plurality of weights;

and

the operation of the of two or more neural network devices comprises learning adaptation of individual ones of the plurality of weights.

19. The method of claim 15, wherein

the configuration of the other neural network device comprises one or more high-level neuromorphic description instructions configured to be executed by the two or more neural network devices;

the observing comprises comparing the one or more high-level neuromorphic description instructions to a template;

the reproducing the configuration is based on the one or more high-level neuromorphic description instructions matching a portion of the template; and

the preventing reproduction is based on the one or more high-level neuromorphic description instructions deviating by an established measure from the template.

20. A cloud server system, comprising:

a storage apparatus configured to store a plurality of neuromorphic apparatus state data;

a network interface configured to receive one or more subscriber requests, the one or more subscriber requests including a first subscriber request; and

one or more processors configured to execute computer program modules, individual ones of the one or more processors being communicatively coupled with the storage apparatus and the network interface, the computer program modules being executable to perform a method to facilitate a business transaction, the method comprising;

authenticating a subscriber accessing the cloud server system via a user interface device;

receiving the first subscriber request from the subscriber, the first subscriber request being for one or more of the plurality of neuromorphic apparatus state data;

determining whether the subscriber is authorized to receive the one or more of the plurality of neuromorphic apparatus state data; and

based on the determination, effectuating the business transaction by transmitting the one or more of the plurality of neuromorphic apparatus state data to one or both of;

(i) a neuromorphic apparatus associated with the subscriber or (ii) the user interface device.