Methods and systems for implementing deep reinforcement module networks for autonomous systems control
A novel architecture for a network of deep reinforcement modules that enables cross-functional and multi-system coordination of autonomous systems for self-optimization with a reduced computational footprint is disclosed. Each deep reinforcement module in the network is comprised of either a single artificial neural network or a deep reinforcement module sub-network. DReMs are designed independently, decoupling each requisite function. Each module of a deep reinforcement module network is trained independently through deep reinforcement learning. By separating the functions into deep reinforcement modules, reward functions can be designed for each individual function, further simplifying the development of a full suite of algorithms while also minimizing training time. Following training, the DReMs are integrated into the full deep reinforcement module network, which is then refined through additional reinforcement training or genetic multi-objective optimization to maximize the overall performance of the network.
1. A computer-implemented method of training a network of neural networks, the method comprising:
training a first neural network based on a first reward function;
training a second neural network based on a second reward function;
training a first deep reinforcement module based on a third reward function, wherein the first deep reinforcement module comprises the first and second neural networks; and
training a second deep reinforcement module based on a fourth reward function, wherein the second deep reinforcement module comprises the first deep reinforcement module and a third neural network.
2. The method of claim 1 , wherein the first neural network executes on a first processor and the second neural network executes on a second processor.
3. The method of claim 1 , wherein the third neural network is trained based on a fifth reward function.
4. The method of claim 1 , wherein the first and second deep reinforcement modules are connected via a communication network.
5. The method of claim 1 , wherein the first neural network is trained to search for items in visual sensor data.
6. The method of claim 5 , wherein the second neural network is trained to track an item in space.
7. The method of claim 6 , wherein the first deep reinforcement module is trained to track one or more items found by the first neural network.
8. A system comprising:
a processor; and
a computer-readable medium storing computer-readable instructions, which when executed by the processor, cause the processor to perform:
training a first neural network based on a first reward function;
training a second neural network based on a second reward function;
training a first deep reinforcement module based on a third reward function, wherein the first deep reinforcement module comprises the first and second neural networks; and
training a second deep reinforcement module based on a fourth reward function, wherein the second deep reinforcement module comprises the first deep reinforcement module and a third neural network.
9. The system of claim 8 , wherein the first neural network executes on a first processor and the second neural network executes on a second processor.
10. The system of claim 8 , wherein the third neural network is trained based on a fifth reward function.
11. The system of claim 8 , wherein the first and second deep reinforcement modules are connected via a communication network.
12. The system of claim 8 , wherein the first neural network is trained to search for items in visual sensor data.
13. The system of claim 12 , wherein the second neural network is trained to track an item in space.
14. The system of claim 13 , wherein the first deep reinforcement module is trained to track one or more items found by the first neural network.
15. A computer program product, comprising:
a non-transitory computer readable storage medium having computer readable program code embodied therewith, the computer readable program code configured when executed by a processor to:
train a first neural network based on a first reward function;
train a second neural network based on a second reward function;
train a first deep reinforcement module based on a third reward function, wherein the first deep reinforcement module comprises the first and second neural networks; and
training a second deep reinforcement module based on a fourth reward function, wherein the second deep reinforcement module comprises the first deep reinforcement module and a third neural network.
16. The computer program product of claim 15 , wherein the first neural network is trained to search for items in visual sensor data.
17. The computer program product of claim 16 , wherein the second neural network is trained to track an item in space and the first deep reinforcement module is trained to track one or more items found by the first neural network.
18. The computer program product of claim 15 , wherein the first neural network executes on a first processor and the second neural network executes on a second processor.
19. The computer program product of claim 15 , wherein the third neural network is trained based on a fifth reward function.
20. The computer program product of claim 15 , wherein the first and second deep reinforcement modules are connected via a communication network.