Reinforcement learning exploration by exploiting past experiences for critical events
A computer-implemented method is provided for reinforcement learning performed by a processor. The method includes obtaining, from an environment, a given experience that includes an action, a state and a reward. The method further includes storing the given experience in an experience buffer responsive to a value of the reward included in the given experience exceeding a first threshold. The method also includes responsive to obtaining another experience having another reward that less than or equal to the first threshold, searching the experience buffer for a candidate experience with a similar state to the other experience and copying the candidate experience into an event buffer. The method additionally includes during exploration, selecting an action to be taken to the environment from the event buffer with a predetermined probability.
1 . A computer-implemented method for data determination for reinforcement learning training performed by a processor, the method comprising:
obtaining, from an environment, a given experience that includes a vehicle action, a state and a reward;
configuring an experience buffer and an event buffer for cooperative model training usage such that the experience buffer and the event buffer are no longer used for training after a pre-defined number of training steps and model-free reinforcement learning is performed;
during training, storing the given experience in the experience buffer responsive to a value of the reward included in the given experience not being below an average award amount for a plurality of experiences by a first threshold amount, while excluding from the experience buffer events where an agent dies corresponding to the value of the reward included in the given experience being below the average award amount by the first threshold amount;
further during training, responsive to obtaining another experience, searching the experience buffer for a candidate experience with a similar state but with a better reward and different vehicle action to the other experience and copying the candidate experience into the event buffer storing events where an agent survives;
during exploration, selecting a vehicle action to be taken to the environment from the event buffer with a predetermined probability; and
performing the vehicle action taken from the event buffer to avoid the vehicle action becoming an event where the agent dies, the vehicle action being a controlling of a motor vehicle to perform a braking or steering action for accident avoidance.
2 . The computer-implemented method of claim 1 , further comprising stopping the selecting of the vehicle action from the event buffer after a pre-defined number of steps during a training stage of the reinforcement learning.
3 . The computer-implemented method of claim 2 , further comprising performing random exploration responsive to said stopping step.
4 . The computer-implemented method of claim 1 , further comprising storing in the experience buffer any experiences previously observed except for the experiences that resulted in a corresponding reward that fails to exceed the first threshold.
5 . The computer-implemented method of claim 1 , wherein the state represents a local state in a low-dimensional space.
6 . The computer-implemented method of claim 1 , wherein the method is applied to the plurality of experiences, wherein the first threshold is used to identify critical events, and wherein any of the plurality of experiences unrelated to the critical events are stored in the experience buffer and any of the plurality of experiences related to the critical events are stored in the event buffer.
7 . The computer-implemented method of claim 1 , further comprising plotting, on a display device, a plurality of experiences in a visualization, each of the plurality of experiences having a respective reward to form a plurality of rewards across the plurality of experiences, wherein said plotting step uses the plurality of rewards as weights for the visualization.
8 . A computer program product for data determination for reinforcement learning training performed by a processor, the computer program product comprising a non-transitory computer readable storage medium having program instructions embodied therewith, the program instructions executable by a computer having the processor to cause the computer to perform a method comprising:
obtaining, from an environment, a given experience that includes a vehicle action, a state and a reward;
configuring an experience buffer and an event buffer for cooperative model training usage such that the experience buffer and the event buffer are no longer used for training after a pre-defined number of training steps and model-free reinforcement learning is performed;
during training, storing the given experience in the experience buffer responsive to a value of the reward included in the given experience not being below an average award amount for a plurality of experiences by a first threshold amount, while excluding from the experience buffer events where an agent dies corresponding to the value of the reward included in the given experience being below the average award amount by the first threshold amount;
further during training, responsive to obtaining another experience, searching the experience buffer for a candidate experience with a similar state but with a better reward and different vehicle action to the other experience and copying the candidate experience into the event buffer storing events where an agent survives;
during exploration, selecting a vehicle action to be taken to the environment from the event buffer with a predetermined probability; and
performing the vehicle action taken from the event buffer to avoid the vehicle action becoming an event where the agent dies, the vehicle action being a controlling of a motor vehicle to perform a braking or steering action for accident avoidance.
9 . The computer program product of claim 8 , wherein the method further comprises stopping the selecting of the vehicle action from the event buffer after a pre-defined number of steps during a training stage of the reinforcement learning.
10 . The computer program product of claim 9 , wherein the method further comprises performing random exploration responsive to said stopping step.
11 . The computer program product of claim 8 , wherein the method further comprises storing in the experience buffer any experiences previously observed except for the experiences that resulted in a corresponding reward that fails to exceed the first threshold.
12 . The computer program product of claim 8 , wherein the state represents a local state in a low-dimensional space.
13 . A computer processing system for data determination for reinforcement learning training, comprising:
a memory for storing program code; and
a processor, operatively coupled to the memory, for running the program code to
obtain, from an environment, a given experience that includes vehicle action, a state and a reward;
configuring an experience buffer and an event buffer for cooperative model training usage such that the experience buffer and the event buffer are no longer used for training after a pre-defined number of training steps and model-free reinforcement learning is performed;
during training, store the given experience in the experience buffer responsive to a value of the reward included in the given experience not being below an average award amount for a plurality of experiences by a first threshold amount, while excluding from the experience buffer events where an agent dies corresponding to the value of the reward included in the given experience being below the average award amount by the first threshold amount;
further during training, responsive to obtaining another experience, search the experience buffer for a candidate experience with a similar state but with a better reward and different vehicle action to the other experience and copying the candidate experience into the event buffer storing events where an agent survives;
during exploration, select vehicle action to be taken to the environment from the event buffer with a predetermined probability; and
perform the vehicle action taken from the event buffer to avoid the vehicle action becoming an event where the agent dies, the vehicle action being a controlling of a motor vehicle to perform a braking or steering action for accident avoidance.
14 . The computer-implemented method of claim 1 , wherein during training, the plurality of experiences and their respective frequencies of occurrence with rewards are stored and used to plot the experiences using respective ones of the rewards as weights for visualization.
15 . The computer-implemented method of claim 1 , wherein during training, a next vehicle action is determined by sampling actions for similar events in the event buffer.