Game theoretic decision making
Aspects related to game theoretic decision making may be implemented utilizing a sensor, a memory, and a processor. The sensor may detect other vehicles and corresponding attributes as an observation. The memory may store instructions. The processor may execute the instructions to perform acts, actions, or steps, such as constructing a search tree based on the observation, an initial belief, and a vehicle identified as a current opponent vehicle, performing a Monte Carlo Tree Search (MCTS) on the search tree based on a planning horizon and a time allowance to determine a desired action from a set of ego-actions, executing, via vehicle systems, the desired action, detecting an updated observation associated with one or more of the other vehicles, identifying the other vehicles to be updated as the current opponent vehicle, and updating a root node of the search tree based on the current opponent vehicle.
1. A system for game theoretic decision making, comprising:
a sensor detecting one or more other vehicles and corresponding attributes as an observation;
a memory storing one or more instructions;
a processor executing one or more of the instructions stored on the memory to perform:
constructing a search tree based on the observation, an initial belief, and a vehicle identified as a current opponent vehicle;
performing a Monte Carlo Tree Search (MCTS) on the search tree based on a planning horizon and a time allowance to determine a desired action from a set of ego-actions;
executing, via one or more vehicle systems, the desired action, wherein the desired action include one or more of a lane change, an acceleration action, a deceleration action, or a stopping action;
detecting, via the sensor, an updated observation associated with one or more of the other vehicles;
identifying one or more of the other vehicles to be updated as the current opponent vehicle; and
updating a root node of the search tree based on the current opponent vehicle.
2. The system for game theoretic decision making of claim 1 , wherein the performing the MCTS on the search tree includes performing a tree rollout based on the root node of the search tree and the planning horizon to expand, simulate, and backpropagate the search tree.
3. The system for game theoretic decision making of claim 2 , wherein the performing the tree rollout includes prioritizing unsampled actions.
4. The system for game theoretic decision making of claim 1 , comprising computing, via the processor, a policy for the current opponent vehicle based on a current belief and a pre-computed Quantal Level-k function.
5. The system for game theoretic decision making of claim 4 , comprising sampling, via the processor, one or more opponent actions from the policy for the current opponent vehicle.
6. The system for game theoretic decision making of claim 1 , comprising assigning, via the processor, a default action to non-opponent vehicles of the one or more other vehicles.
7. The system for game theoretic decision making of claim 2 , wherein the expanding the search tree includes generating a child node for a current node based on an ego-action and an opponent action being non-existent within one or more child nodes of the root node.
8. The system for game theoretic decision making of claim 7 , comprising calculating, via the processor, a reward associated with the child node based on a reward function, a discount factor, and an information gain.
9. The system for game theoretic decision making of claim 8 , wherein the information gain is represented by a difference in entropy between the current node and the child node.
10. The system for game theoretic decision making of claim 8 , wherein the information gain is represented by a difference in entropy between a belief associated with the current node and a belief associated with the child node.
11. A computer-implemented method for game theoretic decision making, comprising:
detecting, via a sensor, one or more other vehicles and corresponding attributes as an observation;
constructing a search tree based on the observation, an initial belief, and a vehicle identified as a current opponent vehicle;
performing a Monte Carlo Tree Search (MCTS) on the search tree based on a planning horizon and a time allowance to determine a desired action from a set of ego-actions;
executing, via one or more vehicle systems, the desired action, wherein the desired action include one or more of a lane change, an acceleration action, a deceleration action, or a stopping action;
detecting, via the sensor, an updated observation associated with one or more of the other vehicles;
identifying one or more of the other vehicles to be updated as the current opponent vehicle; and
updating a root node of the search tree based on the current opponent vehicle.
12. The computer-implemented method for game theoretic decision making of claim 11 , wherein the performing the MCTS on the search tree includes performing a tree rollout based on the root node of the search tree and the planning horizon to expand, simulate, and backpropagate the search tree.
13. The computer-implemented method for game theoretic decision making of claim 12 , wherein the performing the tree rollout includes prioritizing unsampled actions.
14. The computer-implemented method for game theoretic decision making of claim 11 , comprising computing a policy for the current opponent vehicle based on a current belief and a pre-computed Quantal Level-k function.
15. The computer-implemented method for game theoretic decision making of claim 14 , comprising sampling one or more opponent actions from the policy for the current opponent vehicle.
16. The computer-implemented method for game theoretic decision making of claim 11 , comprising assigning a default action to non-opponent vehicles of the one or more other vehicles.
17. A game theoretic decision making vehicle, comprising:
one or more vehicle systems;
a sensor detecting one or more other vehicles and corresponding attributes as an observation;
a memory storing one or more instructions;
a processor executing one or more of the instructions stored on the memory to perform:
constructing a search tree based on the observation, an initial belief, and a vehicle identified as a current opponent vehicle;
performing a Monte Carlo Tree Search (MCTS) on the search tree based on a planning horizon and a time allowance to determine a desired action from a set of ego-actions;
executing, via one or more of the vehicle systems, the desired action, wherein the desired action include one or more of a lane change, an acceleration action, a deceleration action, or a stopping action;
detecting, via the sensor, an updated observation associated with one or more of the other vehicles;
identifying one or more of the other vehicles to be updated as the current opponent vehicle; and
updating a root node of the search tree based on the current opponent vehicle.
18. The game theoretic decision making vehicle of claim 17 , wherein the performing the MCTS on the search tree includes performing a tree rollout based on the root node of the search tree and the planning horizon to expand, simulate, and backpropagate the search tree.
19. The game theoretic decision making vehicle of claim 18 , wherein the performing the tree rollout includes prioritizing unsampled actions.
20. The game theoretic decision making vehicle of claim 17 , comprising computing, via the processor, a policy for the current opponent vehicle based on a current belief and a pre-computed Quantal Level-k function.