Solving based introspection to augment the training of reinforcement learning agents for control and planning on robots and autonomous vehicles
Described is a system for controlling a mobile platform. A neural network that runs on the mobile platform is trained based on a current state of the mobile platform. A Satisfiability Modulo Theories (SMT) solver capable of reasoning over non-linear activation functions is periodically queried to obtain examples of states satisfying specified constraints of the mobile platform. The neural network is then trained on the examples of states. Following training on the examples of states, the neural network selects an action to be performed by the mobile platform in its environment. Finally, the system causes the mobile platform to perform the selected action in its environment.
1. A system for controlling a mobile platform, the system comprising:
the mobile platform having one or more sensors thereon; and
one or more processors and a non-transitory computer-readable medium having executable instructions encoded thereon such that when executed, the one or more processors perform operations of:
determining a current states of the mobile platform via the one or more sensors;
initially training a neural network π that is integrated on the mobile platform, wherein the initial training is based on the current states of the mobile platform;
querying a Satisfiability Modulo Theories (SMT) solver when it is determined that a current increment step is on a query schedule,
wherein the query schedule determines when to query the SMT solver to generate a plurality of examples of states satisfying specified constraints of the mobile platform;
modifying the initial training of the neural network π based on the plurality of examples of states;
following training on the plurality of examples of states, selecting an action to be performed by the mobile platform in its environment,
wherein the action is selected from a probability distribution π(s) over a space of valid actions that the mobile platform can take while in the current states; and
causing the mobile platform to perform the selected action in its environment.
2. The system as set forth in claim 1 , wherein the one or more processors further perform an operation of generating the plurality of examples of states utilizing the SMT solver by implementing a state space quantization algorithm.
3. The system as set forth in claim 2 , wherein the one or more processors further perform an operation of applying at least one query constraint when generating the plurality of examples of states.
4. The system as set forth in claim 1 , wherein the one or more processors further perform operations of:
applying a processing algorithm to the plurality of examples of states, resulting in a set of processed examples of states; and
training the neural network π on the set of processed examples of states.
5. A computer implemented method for controlling a mobile platform, the method comprising an act of:
causing one or more processers to execute instructions encoded on a non-transitory computer-readable medium, such that upon execution, the one or more processors perform operations of:
determining a current states of a mobile platform via one or more sensors on the mobile platform;
initially training a neural network π it that is integrated on the mobile platform, wherein the initial training is based on the current states of the mobile platform;
querying a Satisfiability Modulo Theories (SMT) solver when it is determined that a current increment step is on a query schedule,
wherein the query schedule determines when to query the SMT solver to generate a plurality of examples of states satisfying specified constraints of the mobile platform;
modifying the initial training of the neural network π based on the plurality of examples of states;
following training on the plurality of examples of states, selecting an action to be performed by the mobile platform in its environment,
wherein the action is selected from a probability distribution π(s) over a space of valid actions that the mobile platform can take while in the current states; and
causing the mobile platform to perform the selected action in its environment.
6. The method as set forth in claim 5 , wherein the one or more processors further perform an operation of generating the plurality of examples of states utilizing the SMT solver by implementing a state space quantization algorithm.
7. The method as set forth in claim 6 , wherein the one or more processors further perform an operation of applying at least one query constraint when generating the plurality of examples of states.
8. The method as set forth in claim 5 , wherein the one or more processors further perform operations of:
applying a processing algorithm to the plurality of examples of states, resulting in a set of processed examples of states; and
training the neural network π on the set of processed examples of states.
9. A computer program product for controlling a mobile platform, the computer program product comprising:
computer-readable instructions stored on a non-transitory computer-readable medium that are executable by a computer having one or more processors for causing the processor to perform operations of:
determining a current states of the mobile platform via one or more sensors;
initially training a neural network π that is integrated on the mobile platform, wherein the initial training is based on the current states of the mobile platform;
querying a Satisfiability Modulo Theories (SMT) solver when it is determined that a current increment step is on a query schedule,
wherein the query schedule determines when to query the SMT solver to generate a plurality of examples of states satisfying specified constraints of the mobile platform;
modifying the initial training of the neural network π based on the plurality of examples of states;
following training on the plurality of examples of states, selecting an action to be performed by the mobile platform in its environment,
wherein the action is selected from a probability distribution π(s) over a space of valid actions that the mobile platform can take while in the current states; and
causing the mobile platform to perform the selected action in its environment.
10. The computer program product as set forth in claim 9 , further comprising instructions for causing the one or more processors to further perform an operation of generating the plurality of examples of states utilizing the SMT solver by implementing a state space quantization algorithm.
11. The computer program product as set forth in claim 10 , further comprising instructions for causing the one or more processors to further perform an operation of applying at least one query constraint when generating the plurality of examples of states.
12. The computer program product as set forth in claim 9 , further comprising instructions for causing the one or more processors to further perform operations of:
applying a processing algorithm to the plurality of examples of states, resulting in a set of processed examples of states; and
training the neural network π on the set of processed examples of states.