Lifelong learning with a changing action set
Systems and methods are described for a decision-making process that includes an increasing set of actions, compute a policy function for a Markov decision process (MDP) for the decision-making process, wherein the policy function is computed based on a state conditional function mapping states into an embedding space, an inverse dynamics function mapping state transitions into the embedding space, and an action selection function mapping the elements of the embedding space to actions, identify an additional set of actions in the increasing set of actions, update the inverse dynamics function based at least in part on the additional set of actions, update the policy function based on the updated inverse dynamics function and parameters learned during the computing the policy function, and select an action based on the updated policy function.
1. A method for decision-making, comprising:
identifying a decision-making process that includes an increasing set of actions;
computing a policy function for a Markov decision process (MDP) for the decision-making process, wherein the policy function is computed based on a state conditional function mapping states into an embedding space, an inverse dynamics function mapping state transitions into the embedding space, and an action selection function mapping elements of the embedding space to actions;
identifying an additional set of actions for an agent of the MDP;
updating the inverse dynamics function based at least in part on the additional set of actions;
updating the policy function based on the updated inverse dynamics function;
receiving state information for the agent;
selecting an action based on the updated policy function and the state information; and
transmitting an action recommendation to the agent based on the selected action.
2. The method of claim 1 , wherein:
the policy function comprises a composition of the state conditional function and the action selection function, and updating the policy function incorporates parameters learned during the computing the policy function.
3. The method of claim 1 , further comprising:
updating the state conditional function and the action selection function based at least in part on the additional set of actions, wherein the updated policy function is based at least in part on the updated state conditional function and the updated action selection function.
4. The method of claim 3 , wherein:
the action selection function and the inverse dynamics function are updated in a first phase and the state conditional function is updated in a second phase based at least in part on the updated action selection function.
5. The method of claim 3 , wherein:
the state conditional function, the action selection function and the inverse dynamics function are updated simultaneously.
6. The method of claim 1 , further comprising:
maximizing a lower bound of a loss function, wherein the action selection function and the inverse dynamics function are updated based on the maximization.
7. The method of claim 6 , wherein:
the loss function is based on a divergence of states and actions from a set of observed state transitions.
8. The method of claim 1 , further comprising:
computing a policy gradient, wherein the policy function is computed based on the policy gradient.
9. The method of claim 1 , wherein:
the decision-making process comprises a tool recommendation engine and the increasing set of actions is based at least in part on a changing set of tools.
10. The method of claim 1 , wherein:
the decision-making process comprises selecting an advertisement and the increasing set of actions comprises a set of available ads.
11. The method of claim 1 , wherein:
the decision-making process comprises a robot control mechanism and the increasing set of actions comprises a set of available actuator movements.
12. The method of claim 1 , wherein:
the decision-making process comprises a product recommendation system and the increasing set of actions is based at least in part on product availability.
13. The method of claim 1 , wherein:
the decision-making process comprises a medical recommendation system and the increasing set of actions is based at least in part on a changing set of available medical interventions.
14. An apparatus for reinforcement learning, comprising:
a processor and a memory storing instructions and in electronic communication with the processor, the processor being configured to execute the instructions to:
compute a policy function for a Markov decision process (MDP) model of a decision-making process, wherein the policy function is computed based on a state conditional function mapping states into an embedding space, an inverse dynamics function mapping state transitions into the embedding space, and an action selection function mapping elements of the embedding space to actions;
identify an additional set of actions in an increasing set of actions of the MDP;
update the inverse dynamics function and the action selection function during a first phase based at least in part on the additional set of actions by maximizing a lower bound of a loss function;
update the state conditional function during a second phase following the first phase using a policy gradient; and
select an action using the updated state conditional function and the updated action selection function.
15. The apparatus of claim 14 , wherein:
the policy function comprises a neural network including a composition of the state conditional function and the action selection function.
16. The apparatus of claim 14 , wherein:
the loss function is based on a divergence of states and actions from a set of observed state transitions.
17. The apparatus of claim 14 , the processor being further configured to execute the instructions to:
update a critic function by minimizing a temporal difference error.
18. A non-transitory computer-readable medium storing code for reinforcement learning, the code comprising instructions executable by a processor to:
compute a policy function for a Markov decision process (MDP) for a decision-making process that includes an increasing set of actions;
receive state information for an agent of the decision making process;
transmit a recommended action to the agent based on the policy function and the state information;
receive an additional set of actions from the agent;
update the policy function based at least in part on the additional set of actions and parameters learned during the computing the policy function;
receive subsequent state information for the agent; and
transmit a subsequent recommended action to the agent based on the updated policy function and the subsequent state information.
19. The non-transitory computer-readable medium of claim 18 , the code further comprising instructions executable by the processor to:
update a state conditional function and an action selection function based at least in part on the additional set of actions, wherein the updated policy function is based at least in part on the updated state conditional function and the updated action selection function.
20. The non-transitory computer-readable medium of claim 19 , the code further comprising instructions executable by the processor to:
maximize a lower bound of a loss function, wherein the action selection function and the inverse dynamics function are updated based on the maximization.