Contextual memory bandit for proactive dialogs
A proactive interaction system includes memory which stores a contextual model. The contextual model includes supporting memory storing a representation of each of a set of past observations. Each of the past observations having an observed reward for a respective user for a respective action selected from a set of candidate actions. The contextual model is configured for estimating a reward for each of a current set of candidate actions, based on the stored representations of past observations and a representation of a current user. Each candidate action is associated with a respective action representation. A contextual bandit selects one of the candidate actions, based on the estimated reward for each of the set of candidate actions, to optimize a cumulative reward over a sequence of candidate action selections. An act output component performs a user-detectable act based on the selected one of the candidate actions.
1. A proactive interaction system comprising:
a contextual model stored in system memory, the contextual model including supporting memory, the supporting memory storing, as memory cells of the supporting memory, a respective vectorial representation of each of a set of past observations, wherein for each of the set of past observations, there is an observed reward for a respective user for a respective candidate action selected from an identified set of candidate actions, the contextual model configured for estimating a reward for each of a current set of candidate actions based on the stored vectorial representations of past observations and a representation of a current user, each of the candidate actions being associated with a respective action representation;
instructions stored in system memory including:
a contextual bandit which selects one of the current set of candidate actions based on the estimated reward for each of the current set of candidate actions to optimize a cumulative reward over a sequence of candidate action selections;
an act output component which outputs an agent act based on the selected one of the current set of candidate actions, the act output component including components which generate an agent utterance in a natural language as the agent act, unless the selected candidate action is for no action;
a user act input component which detects a user act in the form of user utterance responsive to the agent act;
the proactive interaction system being configured to advance a dialog with the user by generating further agent acts, based on the user utterance; and
a hardware processor which implements the instructions.
2. The system of claim 1 , further comprising at least one of:
a reward computation component which computes an observed reward for the selected action based on an observed user response to the agent act; and
an observation component which updates the supporting memory with a vectorial representation of a new observation for the user and selected action, based on the reward.
3. The system of claim 2 , wherein the observed user response is an utterance.
4. The system of claim 1 , wherein the contextual model includes a memory-enhanced neural network.
5. The system of claim 4 , wherein the memory-enhanced neural network includes an attention mechanism.
6. The system of claim 1 , wherein the supporting memory comprises sets of input and output memory cells that are generated from respective observations with respective transformations.
7. The system of claim 1 , wherein the vectorial representations of each of a set of past observations are each derived from both a representation of a user and a representation of an action.
8. The system of claim 1 , wherein the instructions further comprise a user information component which generates the representation of the user.
9. The system of claim 1 , wherein the instructions further comprise an action identification component which identifies a respective set of at least two candidate actions for each of a plurality of times, from which the one of the candidate actions is selected.
10. The system of claim 1 , wherein the instructions further comprise a learning component which updates parameters of the model based on the estimated reward and observed reward.
11. The system of claim 1 , wherein the agent actions are related to points of interest and events.
12. The system of claim 1 , wherein the system, for each of a plurality of times, selects a candidate action and outputs a user-detectable act based on the selected candidate action.
13. The system of claim 1 , wherein the set of candidate actions includes at least two candidate actions.
14. The system of claim 13 , wherein the set of candidate actions includes a candidate action corresponding to no action and wherein when the selected candidate action is the candidate action corresponding to no action, the agent act corresponds to no act being output.
15. A proactive interaction method comprising:
providing, in system memory, a contextual model, the contextual model including supporting memory, the supporting memory including, for each of at least two hops, input and output supporting memories, storing respective input and output memory cells therein, the memory cells storing input and output representations of each of a set of past observations for which there is an observed reward for a respective user for a respective action selected from a set of candidate actions, the input and output representations being obtained by transforming each of the set of past observations with respective embedding matrices, the contextual model configured for estimating a reward for each of a set of candidate actions based on the stored input and output representations of past observations and a representation of a current user, each candidate action being associated with a respective action representation;
for a sequence of iterations, with a hardware processor:
for a current user, with the model, computing an expected reward for each of a current set of at least two candidate actions; and
for at least one iteration in the sequence of iterations:
sampling one of the candidate actions based on the estimated reward for each of the set of candidate actions to optimize a cumulative reward over a sequence of candidate action selections,
outputting an agent act, based on the selected one of the candidate actions, the agent act comprising an utterance in a natural language,
computing a reward for the selected action based on an observed user response to the agent act, and
updating the supporting memory with input and output representations of a new observation for the user and selected action, based on the reward, and
advancing a dialog with the user by generating further agent acts, based on the observed user response to the agent act.
16. The method of claim 15 , wherein for at least one of the sequence of iterations, the selected action is no action and no agent act is output.
17. The method of claim 15 , wherein the computing of the expected reward, sampling one of the candidate actions, computing a reward for the selected action, and updating the supporting memory are performed with a processor.
18. A vehicle comprising the proactive interaction system of claim 1 .