Reinforcement learning using adaptive return calculation method

The system addresses inefficiencies in training machine learning models for agent control by using extrinsic and intrinsic rewards with dynamic return calculation schemes, enhancing performance and reducing resource consumption.

JP7581358B2Active Publication Date: 2024-11-12ジーディーエム·ホールディング·エルエルシー
View PDF 1 Cites 0 Cited by

Patent Information

Application Number
JP2022548005
Authority / Receiving Office
JP · JP
Patent Type
Patents
Current Assignee / Owner
Priority Date
2020-02-07
Filing Date
2021-02-08
Publication Date
2024-11-12
Estimated Expiration
2041-02-08

AI Technical Summary

Technical Problem

Existing machine learning models for agent control in environments face inefficiencies in training, particularly in balancing exploration and exploitation, leading to suboptimal performance and increased computational resource consumption.

Method used

A system that trains action selection neural networks using a combination of extrinsic and intrinsic rewards, with an adaptation mechanism to dynamically adjust return calculation schemes, employing two separate neural networks for improved stability and efficiency.

Benefits of technology

The system achieves improved performance and reduced computational resources by optimizing the balance between exploration and exploitation, resulting in more effective agent control across various tasks.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 0007581358000013
    Figure 0007581358000013
  • Figure 0007581358000014
    Figure 0007581358000014
  • Figure 0007581358000015
    Figure 0007581358000015
Patent Text Reader

Abstract

A method, system, and apparatus, including a computer program encoded on a computer storage medium, for reinforcement learning using an adaptive return computation scheme. In one aspect, the method includes maintaining data defining a policy for selecting among a plurality of different return computation schemes, each return computation scheme assigning a different importance to exploring an environment while performing an episode of a task, selecting a return computation scheme from the plurality of different return computation schemes using the policy, controlling an agent to perform the episode of the task to maximize a return computed according to the selected return computation scheme, identifying a reward generated as a result of the agent performing the episode of the task, and updating the policy to select among the plurality of different return computation schemes using the identified reward.
Need to check novelty before this filing date? Find Prior Art

Description

[Technical field]

[0001] This specification relates to processing data using machine learning models. [Background technology]

[0002] A machine learning model receives input and generates an output, e.g., a predicted output, based on the received input. Some machine learning models are parametric models and generate an output based on the received input and the values ​​of the parameters of the model.

[0003] Some machine learning models are deep models that employ multiple layers of models to generate an output for a received input. For example, a deep neural network is a deep machine learning model that includes an output layer and one or more hidden layers that each apply a nonlinear transformation to the received input to generate an output. [Prior art documents] [Non-patent literature]

[0004] [Non-Patent Document 1] T. Schul et al., "Prioritized experience replay," arXiv:1511.05952v4 (2016) [Non-Patent Document 2] Y. Burda et al., "Exploration by random network distillation," arXiv:1810.12894v1, (2018) Summary of the Invention [Means for solving the problem]

[0005] This specification describes a system, implemented as a computer program on one or more computers at one or more locations, for training a set of one or more action selection neural networks to control an agent interacting with an environment.

[0006] Particular embodiments of the subject matter described in this specification can be implemented to realize one or more of the following advantages.

[0007] The system described herein maintains a policy that the system uses to select an optimal return calculation scheme to use for any given task episode performed during training of a set of one or more action selection neural networks used to control an agent interacting with an environment to perform a task. Each possible return calculation scheme assigns a different importance to the exploration of the environment while the agent interacts with the environment. In particular, each possible return compensation scheme specifies a discount factor used to calculate the return, an intrinsic reward scaling factor used to calculate the total reward, or both. More specifically, the system uses an adaptation mechanism to adjust the policy throughout the training process, resulting in a greater likelihood that different return calculation schemes will be selected at different points during training. The use of this adaptation mechanism allows the system to effectively select, at any given time during training, the most appropriate time horizon for calculating the return, the most appropriate degree of exploration, or both. This results in a trained neural network that can exhibit improved performance when controlling an agent to perform any of a variety of tasks.

[0008] In addition, in some implementations, the system uses two separate neural networks to estimate the "value" of any given action: an extrinsic return action selection neural network trained using only extrinsic rewards, and an intrinsic return action selection neural network trained using only intrinsic rewards. This novel parameterization of the architecture allows for more consistent and stable learning, reduces the training time and amount of computational resources required for training, and improves the performance of the resulting trained neural network. In particular, these two neural networks can be trained on the same trajectory (but using trajectories and different types of rewards), thus achieving improved consistency and stability without a loss of data efficiency.

[0009] Compared to conventional systems, the systems described herein may consume fewer computational resources (e.g., memory and computing power) by training action selection neural networks to achieve an acceptable level of performance for fewer training iterations. Moreover, a set of one or more action selection neural networks trained by the systems described herein may select actions that enable an agent to complete a task more effectively (e.g., more quickly) than action selection neural networks trained by alternative systems.

[0010] The details of one or more embodiments of the subject matter herein are set forth in the accompanying drawings and the description below. Other features, aspects, and advantages of the subject matter will become apparent from the description, drawings, and claims. [Brief description of the drawings]

[0011] [Figure 1] FIG. 1 illustrates an exemplary action selection system. [Diagram 2] FIG. 1 illustrates an exemplary training system. [Diagram 3]FIG. 1 illustrates an exemplary intrinsic reward system. [Figure 4] 1 is a flow diagram of an example process for controlling an agent to perform a task episode and for updating a return calculation method selection policy. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0012] Like reference numbers and designations in the various drawings indicate like elements.

[0013] 1 illustrates an exemplary action selection system 100. Action selection system 100 is one example of a system implemented as a computer program on one or more computers at one or more locations in which the systems, components, and techniques described below may be implemented.

[0014] The action selection system 100 controls an agent 104 that interacts with an environment 106 to complete a task by using one or more action selection neural networks 102 and policy data 120 to select an action 108 to be performed by the agent 104 at each of a number of time steps during execution of an episode of the task.

[0015] An "episode" of a task is a sequence of iterations during which the agent attempts to execute an instance of the task starting from some starting state of the environment. In other words, each task episode begins with the environment in an initial state, e.g., a fixed or randomly selected initial state, and ends when the agent successfully completes the task or when some termination criterion is met, e.g., the environment enters a state designated as a terminal state, or the agent performs a threshold number of actions without successfully completing the task.

[0016] At each time step during any given task episode, the system 100 receives observations 110 that characterize the current state of the environment 106 at the time step, and accordingly selects an action 108 to be performed by the agent 104 at that time step. After the agent performs an action 108, the environment 106 transitions to a new state and the system 100 receives an extrinsic reward 130 from the environment 106.

[0017] In general, the extrinsic reward 130 is a scalar, numerical value that characterizes the agent's 104 progress toward completing a task.

[0018] As a particular example, the extrinsic reward 130 may be a sparse binary reward that is zero if the task is not successfully completed and one if the task is successfully completed as a result of the action being performed.

[0019] As another particular example, the extrinsic reward 130 may be a dense reward that measures the agent's progress toward completing the task as determined based on individual observations received during an episode of attempting to perform the task, i.e., whereby a non-zero reward may be received frequently before the task is successfully completed.

[0020] While performing any given task episode, the system 100 selects actions to attempt to maximize the returns received over the task episode.

[0021] That is, at each time step during the episode, the system 100 selects an action that attempts to maximize the return received for the remainder of the task episode starting at the time step.

[0022] In general, at any given time step, the return received is a combination of the rewards received at time steps following the given time step in the episode.

[0023] For example, at time step t, the return is Σ i Gamma i-t-1 r i , where i varies over all time steps after t in the episode or for some fixed number of time steps after t in the episode, γ is a discount factor, and r i is the total reward at time step i. As can be seen from the above equation, a higher value of the discount factor results in a longer time horizon for the return calculation, i.e., rewards from time steps that are further away in time from time step t are given more weight in the return calculation.

[0024] In some implementations, the total reward for a given time step is equal to the extrinsic reward received in the time step, i.e., received as a result of actions performed in previous time steps.

[0025] In some other implementations, the system 100 also obtains, i.e., receives or generates, an intrinsic reward 132 from the observations received at least in the time step. The intrinsic reward 132 characterizes the agent's progress toward exploring the environment at the time step in a manner independent of the task performed, i.e., rather than characterizing the agent's progress toward completing the task at the time step, as the extrinsic reward 130 does. That is, the intrinsic reward 132 measures the exploration of the environment rather than measuring the agent's performance on the task. For example, the intrinsic reward may be a value indicative of the extent to which the observations provide information about the entire environment and / or possible configurations of objects within the environment, e.g., if the environment is a real-world environment and the observations are images (or other sensor data) of a corresponding portion of the environment, the intrinsic reward may be a value indicative of how much of the environment appeared in at least one of the images. The calculation of the intrinsic reward for a time step is described in more detail below with reference to FIG. 2 and FIG. 3.

[0026] In these implementations, the system 100 may determine a “total” reward received by the agent in a time step based on at least (i) the extrinsic reward for the time step, (ii) the intrinsic reward for the time step, and (iii) an intrinsic reward scaling factor.

[0027] As a particular example, the system 100 calculates a total reward γ t For example,

number

number

number

[0028] Policy data 120 is data that defines a policy for selecting between different return calculation schemes from a set of return calculation schemes.

[0029] Each return computation scheme in the set assigns a different importance to exploring the environment while performing an episode of the task. In other words, some return computation schemes in the set assign a greater importance to exploring the environment, i.e., gathering new information about the environment, while other return computation schemes in the set assign a greater importance to exploiting the environment, i.e., exploiting current knowledge about the environment.

[0030] As one particular example, each return calculation scheme may define at least one respective discount factor γ used in combining rewards to generate a return. In other words, some return calculation schemes may define relatively larger discount factors, i.e., discount factors that result in rewards in future time steps are weighted relatively more heavily in the return calculation for the current time step than other return calculation schemes.

[0031] As another particular example, each return calculation scheme may define at least one respective intrinsic reward scaling factor β that defines the importance of intrinsic rewards relative to extrinsic rewards received from the environment when generating return.

[0032] In particular, as explained above, the intrinsic reward scaling factor defines the extent to which the intrinsic reward for a given time step is scaled before being added to the extrinsic reward for a given time step to generate the total reward for the time step. In other words, some return calculation schemes may define a relatively large intrinsic reward scaling factor, i.e., a scaling factor that results in the intrinsic reward at a time step being assigned a relatively larger weight in the calculation of the total return for the time step than other return calculation schemes.

[0033] As another specific example, each return calculation scheme may define a respective discount factor-intrinsic reward scaling factor pair, i.e., γ-β pairs, whereby each scheme in the set defines a different combination of values ​​for discount factors and scaling factors from each other set in the scheme.

[0034] Prior to running an episode, system 100 selects a return calculation scheme from a number of different schemes in the set using the policy currently specified by policy data 120. For example, system 100 may select a scheme based on a reward score assigned to the scheme by the policy. Selecting a scheme is described in more detail below with reference to FIG. 4. As described in more detail below, the policy is adaptively modified during training such that different schemes are more likely to be selected at different times during training.

[0035] The system 100 then controls the agent 104 to execute the task episode according to the selected strategy, i.e., using one or more action selection neural networks 102 to maximize the return calculated using the selected strategy.

[0036] To do so, at each time step within the episode, the system 100 uses an action selection neural network 102 to process inputs including: (i) observations 110 characterizing the current state of the environment at the time step, and (ii) selected scheme data that specifies a return computation scheme 112 selected to generate an action score 114. The selected scheme data may specify the scheme 112 to be selected in any suitable manner, for example, as a learned embedding or as a one-hot vector having a value of "1" in the position corresponding to the selected scheme and a value of "0" in the positions corresponding to all other schemes in the set.

[0037] The action scores 114 (also called "return estimates") may include respective numerical values ​​for each action in a set of possible actions and are used by the system 100 to select an action 108 to be performed by the agent 104 at a time step.

[0038] The action selection neural network 102 may be understood as implementing a family of action selection policies indexed by the possible return computation schemes in the set. In particular, the training system 200 (described in more detail with reference to FIG. 2) may train the action selection neural network 102 such that a selected return computation scheme characterizes the degree to which the corresponding action selection policy is "exploratory," i.e., to select actions that cause the agent to explore the environment. In other words, the training system 200 trains the action selection neural network 102 such that tuning the neural network on the selected scheme causes the network to generate outputs that define an action selection policy that places more or less emphasis on exploration versus exploitation of the environment depending on which scheme is selected.

[0039] In some implementations, the one or more action selection neural networks 102 are a single neural network that receives inputs including the observations 110 and data defining the selected return calculation scheme 112 and produces as output the action scores 114. That is, in these implementations, the action scores 114 are the output of the single action selection neural network 102.

[0040] The action selection neural network 102 may be implemented using any suitable neural network architecture that enables performance of its described functions. In one example, the action selection neural network 102 may include an "embedding" sub-network, a "core" sub-network, and a "selection" sub-network. A sub-network of a neural network refers to a group of one or more neural network layers in a neural network. When the observations are images, the embedding sub-network may be a convolutional sub-network, i.e., it includes one or more convolutional neural network layers configured to process the observations for a time step. When the observations are lower dimensional data, the embedding sub-network may be a fully connected sub-network. The core sub-network may be, for example, a recurrent sub-network including one or more long short-term memory (LSTM) neural network layers configured to process (i) the output of the embedding sub-network, (ii) selected scheme data representing a selected return calculation scheme, and optionally (iii) data specifying a most recently received extrinsic reward (and optionally an intrinsic reward) and / or a most recently performed action. The selection sub-network may be configured to process the output of the core sub-network to generate an action score 114.

[0041] In some other implementations, the one or more action selection neural networks 102 are two separate neural networks: (i) an intrinsic reward action selection neural network that estimates an intrinsic return calculated solely from intrinsic rewards generated by the intrinsic reward system based on observations received during interaction with the environment, and (ii) an extrinsic reward action selection neural network that estimates an extrinsic return calculated solely from extrinsic rewards received from the environment as a result of interaction with the environment.

[0042] In these implementations, the two separate neural networks may have the same architecture but may have different parameter values ​​as a result of being trained to estimate different quantities, i.e., the intrinsic reward action selection neural network is trained to estimate intrinsic returns and the extrinsic reward action selection neural network is trained to estimate extrinsic returns.

[0043] In these implementations, the system 100 processes the inputs using an intrinsic reward action selection neural network to generate a respective intrinsic action score for each action (an "estimated intrinsic return"), and processes the inputs using an extrinsic reward action selection neural network to generate a respective extrinsic action score for each action (an "estimated extrinsic return").

[0044] The system 100 may then combine the intrinsic and extrinsic action scores for each action according to the intrinsic reward scaling factor to generate a final action score for the action (a "final return estimate").

[0045] As an example, suppose the jth strategy is selected, the final action score Q(x, a, j: θ) for action a in response to observation x is Q(x, a, j: θ)=Q(x, a, j: θ e )+β j Q(x, a, j: θ i ) where Q(x, a, j: θ e ) is the exogenous action score for action a, and Q(x, a, j: θ i ) is the intrinsic action score for action a, and β j is the scaling factor within the jth scheme.

[0046] As another example, the final action score Q(x, a, j: θ) is Q(x, a, j: θ)=h(h -1 (Q(x, a, j: θ e ))+β j h -1 (Q(x, a, j: θ i ))) where h is a monotonically increasing, invertible squashing function that scales the state-action value functions, i.e., the extrinsic and intrinsic reward functions, to make them easy to estimate for neural networks.

[0047] Therefore, β in the return method j Different values ​​of x cause the predictions of the intrinsic action selection neural network to be weighted differently when computing the final action score.

[0048] The system 100 can use the action scores 114 to select an action 108 to be performed by the agent 104 at a time step. For example, the system 100 can select an action 108 to be performed by the agent 104 by processing the action scores 114 to generate a probability distribution over a set of possible actions and then sampling the actions according to the probability distribution. The system 100 can generate a probability distribution over the set of possible actions by processing the action scores 114, for example, using a soft-max function. As another example, the system 100 can select the action 108 to be performed by the agent 104 as the possible action associated with the highest action score 114. Optionally, the system 100 can select an action 108 to be performed by the agent 104 at a time step using an exploration policy, for example, an ε-greedy exploration policy in which the system 100 selects the action with the highest final return estimate with probability 1-ε and selects a random action from the set of actions with probability ε.

[0049] Once a task episode is executed, i.e., once the agent successfully executes the task or some termination criteria for the task episode are met, the system 100 can use the results of the task episode to (i) update the policy for selecting between the return computation schemes currently specified by the policy data 120, (ii) train the action selection neural network 102, or both.

[0050] More generally, both the policy defined by the policy data 120 and the parameter values ​​of the action selection neural network 102 are updated during training based on trajectories generated as a result of the agent's 104 interactions with the environment.

[0051] In particular, during training, the system 100 updates the return scheme selection policy using recently generated trajectories, e.g., trajectories generated within a fixed-size sliding window of the most recently generated trajectories. By using this adaptation mechanism to adjust the return computation scheme selection policy throughout the training process, different return computation schemes are more likely to be selected at different points during the training of the action selection neural network. Using this adaptation mechanism allows the system 100 to effectively select the most appropriate time horizon, the most appropriate degree of exploration, or both, at any given time during training. This results in a trained neural network that can exhibit improved performance when controlling an agent to perform any of a variety of tasks.

[0052] Updating the return scheme selection policy is described in more detail below with reference to FIG.

[0053] The system 100 trains the action selection neural network 102 using trajectories that are generated by the system 100 and added to a data store called a replay buffer. In other words, at defined intervals during training, the system 100 samples trajectories from the replay buffer and uses the trajectories to train the neural network 102. In some implementations, trajectories used to update the return strategy selection policy are also added to the replay buffer for later use in training the neural network 102. In other implementations, trajectories used to update the return strategy selection policy are not added to the replay buffer and are only used to update the policy. For example, the system 100 may alternate between executing task episodes used to update the policy and executing task episodes that are added to the replay buffer.

[0054] Training the action selection neural network is described below with reference to FIG.

[0055] In some implementations, the environment is a real-world environment and the agent is a mechanical agent that interacts with the real-world environment, e.g., by moving through and / or rotating within the environment and / or by changing its configuration, and / or by modifying the real-world environment. For example, the agent may be a robot that interacts with the environment, e.g., to locate an object of interest in the environment, move an object of interest to a particular location within the environment, physically manipulate an object of interest in the environment, or navigate to a particular destination within the environment, or the agent may be an autonomous or semi-autonomous ground, air, or sea (underwater) vehicle that moves through the environment toward a specified destination within the environment.

[0056] In these implementations, the observations may include, for example, one or more of images, object position data, and sensor data, e.g., from image, distance or position sensors or from actuators, to capture the observations as the agent interacts with the environment.

[0057] For example, in the case of a robot, the observations may include data characterizing the current state of the robot, such as one or more of joint positions, joint velocities, joint forces, torques or accelerations, such as gravity compensated torque feedback, and the global or relative pose of an item being grasped by the robot.

[0058] In the case of a robot or other mechanical agent or vehicle, the observables may similarly include one or more of the position, linear or angular velocity, force, torque or acceleration, and global or relative pose of one or more parts of the agent. The observables may be defined in one, two or three dimensions and may be absolute and / or relative observables.

[0059] The observations may include, for example, data acquired by one or more sensor devices that sense the real-world environment, e.g., sensed electrical signals such as motor currents or temperature signals, and / or image or video data, e.g., from a camera or LIDAR sensor, e.g., data from a sensor on the agent or from a sensor located remotely from the agent in the environment.

[0060] In the case of electronic agents, the observations may include data from one or more sensors monitoring parts of a plant or service facility, such as current, voltage, power, temperature, and other sensors and / or electronic signals that represent the functioning of electronic and / or mechanical items of equipment.

[0061] An action may be a control input for controlling a robot, e.g., torque or higher level control commands for the joints of the robot, or an autonomous or semi-autonomous ground or air or sea (subsea) vehicle, e.g., torque or higher level control commands for a control surface or other control element of the vehicle.

[0062] In other words, an action may include, for example, position, velocity, or force / torque / acceleration data for one or more joints of a robot or part of another mechanical agent. Actions may additionally or alternatively include electronic control data, such as motor control data, or more generally, data for controlling one or more electronic devices in the environment, the control of which affects the observed state of the environment. For example, in the case of an autonomous or semi-autonomous ground, air, or sea (undersea) vehicle, actions may include actions for controlling navigation, e.g., steering, and motion, e.g., braking and / or acceleration of the vehicle.

[0063] In some implementations, the environment is a simulated environment and the agent is implemented as one or more computers that interact with the simulated environment.

[0064] For example, the simulated environment may be a simulation of a robot or vehicle, and the action selection network may be trained on the simulation. For example, the simulated environment may be a motion simulation environment, such as a driving simulation or a flight simulation, and the agent may be a simulated vehicle that navigates through the motion simulation. In these implementations, the actions may be control inputs for controlling a simulated user or a simulated vehicle.

[0065] In another example, the simulated environment may be a video game and the agent may be a simulated user playing the video game.

[0066] In a further example, the environment may be a protein folding environment where each state is a respective state of a protein chain and the agent is a computer system for determining how to fold the protein chain. In this example, the actions are possible folding actions for folding the protein chain and the results to be achieved may include, for example, folding the protein so that it is stable and so that it achieves a particular biological function. As another example, the agent may be a mechanical agent that performs or controls protein folding actions automatically selected by the system without human interaction. The observations may include direct or indirect observations of the protein's state and / or may be derived from a simulation.

[0067] In general, for a simulated environment, the observations may include simulated versions of one or more of the observations or types of observations previously described, and the actions may include simulated versions of one or more of the actions or types of actions previously described.

[0068] Training an agent in a simulated environment may allow the agent to learn from large amounts of simulated training data while avoiding risks associated with training an agent in a real-world environment, such as damage to the agent due to the execution of poorly selected actions. Agents trained in a simulated environment may then be deployed into a real-world environment.

[0069] In some other applications, the agent may control actions in a real-world environment, including, for example, items of equipment in a data center, or a power or water distribution system drawn from a grid, or in a manufacturing plant or service facility. The observations may then relate to the operation of the plant or facility. For example, the observations may include observations of power or water usage by equipment, or observations of control of power generation or distribution, or observations of resource usage or waste production. The agent may control actions in the environment to improve efficiency, for example, by reducing resource usage, and / or to reduce the environmental impact of operations in the environment, for example, by reducing waste. The actions may include actions that control or impose operating conditions on items of equipment in the plant / facility, and / or actions that effect changes in settings in the operation of the plant / facility, for example, to adjust or turn on / off components of the plant / facility.

[0070] Optionally, in any of the above implementations, the observations at any given time step may include data from previous time steps that may be useful in characterizing the environment, such as actions performed in the previous time step, rewards received in response to being performed in the previous time step, etc.

[0071] The training system 200 may determine the reward received by the agent 104 at each time step and may train the action selection neural network 102 using reinforcement learning techniques to optimize a cumulative measure of the reward received by the agent. The reward received by the agent may be represented, for example, by a scalar numerical value. The training system 200 may determine the reward received by the agent at each time step based at least in part on the intrinsic reward scaling factor 112 processed by the action selection neural network 102 at the time step. In particular, the value of the intrinsic reward scaling factor 112 may determine the degree to which exploration of the environment 106 contributes to the reward received by the agent. In this manner, the training system 200 may train the action selection neural network 102 such that for higher values ​​of the intrinsic reward scaling factor 112, the action selection neural network 102 selects actions that cause the agent to explore the environment more quickly.

[0072] 2 illustrates an exemplary training system 200. Training system 200 is one example of a system implemented as a computer program on one or more computers at one or more locations in which the systems, components, and techniques described below are implemented.

[0073] The training system 200 is configured to train the action selection neural network 102 (as described with reference to FIG. 1 ) to optimize a cumulative measure of total reward received by the agent by executing actions selected using the action selection neural network 102.

[0074] As explained above, the training system 200 can determine the “total” reward 202 received by the agent in a time step based on at least (i) the “extrinsic” reward 204 for the time step, (ii) the “exploration” reward 206 for the time step, and (iii) the intrinsic reward scaling factor defined by the sampled return computation scheme 210 for the episode to which the time step belongs.

[0075] As described above, the intrinsic reward 206 may characterize the agent's progress toward exploring the environment at the time step. For example, the training system 200 may determine the intrinsic reward 206 for the time step based on a similarity measure between (i) an embedding of the observation 212 characterizing the state of the environment at the time step and (ii) an embedding of one or more previous observations characterizing the state of the environment at the respective previous time step. In particular, a lower similarity between the embedding of the observation at the time step and the embedding of the observation at the previous time step may indicate that the agent is exploring aspects of the environment that have not been seen before, thus resulting in a higher intrinsic reward 206. The training system 200 may generate the intrinsic reward 206 for the time step by processing the observations 212 characterizing the state of the environment at the time step using an intrinsic reward system 300, which will be described in more detail with reference to FIG. 3.

[0076] To train the action selection neural network 102, the training system 200 obtains a "trajectory" that characterizes the agent's interactions with the environment over one or more (successive) time steps during a task episode. In particular, the trajectory may specify, for each time step, (i) an observation 212 that characterizes the state of the environment at the time step, (ii) an intrinsic reward 202 for the time step, and (iii) an extrinsic reward for the time step. The trajectory also specifies a return computation scheme that was used to select the action corresponding to the trajectory, i.e., to be performed by the agent during the task episode.

[0077] When a single action selection neural network is present, the training engine 208 can then train the action selection neural network 102 by calculating the respective total rewards for each time step from the intrinsic rewards for the time step, from the extrinsic rewards for the time step, and from the intrinsic reward scaling factors described above, i.e., either a constant intrinsic reward scaling factor or an intrinsic reward scaling factor defined by a return computation scheme corresponding to the trajectory if different return computation schemes define different intrinsic reward scaling factors.

[0078] The training engine 208 can then train the action selection neural network on the trajectory using a reinforcement learning technique, which can be, for example, a Q-learning technique, such as a retrace Q-learning technique, or a retrace Q-learning technique with a transformed Bellman operator, such that the action selection neural network is a Q-neural network and the action scores are Q-values ​​that estimate the expected returns received if the corresponding actions are performed by the agent.

[0079] In general, reinforcement learning techniques use discount factors for rewards received at future time steps in a trajectory, estimate future returns received after any given time step, or both, to calculate a target output for training an action selection neural network. When training on a trajectory, the system 200 uses the discount factors in the return calculation scheme corresponding to the trajectory. This results in the action selection neural network being trained to generate action scores that weight future rewards differently when adjusted on different discount factors.

[0080] When there are two action selection neural networks 102, the training engine 208 can train the two action selection neural networks 102 separately on the same trajectory, i.e., without calculating the total reward for the time steps in the trajectory. More specifically, the training engine 208 can train the intrinsic reward action selection neural network using reinforcement learning techniques, but using only the intrinsic reward for the time steps in the trajectory. The training engine 208 can also train the extrinsic reward action selection neural network using reinforcement learning techniques, but using only the extrinsic reward for the time steps in the trajectory.

[0081] To train both neural networks, the training engine 208 can use the same "goal" policy for training, i.e., a policy that selects actions that maximize the total reward calculated using the scaling factors in the corresponding return calculation scheme 210. Similarly, as explained above, when training on a trajectory, the system 200 uses the discount factors in the return calculation scheme corresponding to the trajectory. This results in the action selection neural networks being trained to generate action scores that weight future rewards differently when adjusted for different discount factors.

[0082] In doing so, system 200 trains an extrinsic reward action selection neural network to generate an extrinsic action score (an "estimated extrinsic return") that is an estimate of the intrinsic return that would be received if the corresponding action were performed by the agent, while training an intrinsic reward action selection neural network to generate an intrinsic action score (an "estimated intrinsic return") that is an estimate of the intrinsic return that would be received if the corresponding action were performed by the agent.

[0083] In some implementations, during training, the system 100 can generate trajectories for use by the training system 200 in training the action selection neural network 102 using multiple actor computing units. In some of these implementations, each actor computing unit maintains and separately updates a policy for selecting between multiple different return computation schemes. This can be advantageous when different actor computing units use different values ​​of ε in ε-greedy control, or in some cases, control agents differently. In some other implementations, the system can maintain a central policy that is the same for all of the actor computing units.

[0084] When multiple actor computing units are used, each actor computing unit may iteratively perform the operations described above with reference to FIG. 1 to control an instance of an agent to execute task episodes and use the results of the agent's interactions during the task episodes to update a return computation scheme selection policy, either a central policy or a policy maintained separately by the actor computing unit, and to generate training data for training an action selection neural network.

[0085] A computing unit may be, for example, a computer, a core in a computer with multiple cores, or other hardware or software, for example, a dedicated thread in a computer that can perform operations separately. A computing unit may include a processor core, a processor, a microprocessor, a dedicated logic circuit, such as an FPGA (field programmable gate array) or an ASIC (application specific integrated circuit), or any other suitable computing unit. In some examples, the computing units are all the same type of computing unit. In other examples, the computing units may be different types of computing units. For example, one computing unit may be a CPU while another computing unit may be a GPU.

[0086] The training system 200 stores the trajectories generated by each actor computing unit in a data store called a replay buffer, and in each of multiple training iterations, samples a set of trajectories from the replay buffer for use in training the action selection neural network 102. The training system 200 may sample trajectories from the replay buffer according to a prioritized experience replay algorithm, for example, by assigning a respective score to each stored trajectory and sampling the trajectory associated with that score. An exemplary prioritized experience replay algorithm is described in T. Schul et al., "Prioritized experience replay," arXiv:1511.05952v4 (2016).

[0087] Possible intrinsic reward scaling factors included within the return calculation formulas in the set

number

[0088] Other possible intrinsic reward scaling factors can be respective positive numbers (generally all different from each other) and can be viewed as having the action selection neural network implement respective "exploratory" action selection policies that encourage the agent to explore the environment as well as to solve its task to the extent dictated by the corresponding intrinsic reward scaling factor.

[0089] The action selection neural network 102 can use information provided by the exploration action selection policy to learn a more effective exploitation action selection policy. The information provided by the exploration policy may include, for example, information stored in the shared weights of the action selection neural network. By jointly learning a range of action selection policies, the training system 200 may enable the action selection neural network 102 to learn each individual action selection policy more efficiently, for example, over fewer training iterations. Moreover, learning the exploration policy allows the system to continue training the action selection neural network even when the extrinsic reward is sparse, for example, rarely non-zero.

[0090] After training of the action selection neural network is complete, the system 100 can either update the strategy selection policy and continue to select strategies as described above, or fix the strategy selection policy and control the agent by greedily selecting the strategy for which the strategy selection policy shows the highest reward score.

[0091] In general, the system can use any type of reward that characterizes exploration progress rather than task progress as an intrinsic reward. A specific example of an intrinsic reward and a description of a system for generating intrinsic rewards is described below with reference to FIG.

[0092] 3 illustrates an exemplary intrinsic reward system 300. The intrinsic reward system 300 is one example of a system implemented as a computer program on one or more computers at one or more locations where the systems, components and techniques described below are implemented.

[0093] The intrinsic reward system 300 is configured to process current observations 212 that characterize the current state of the environment to generate intrinsic rewards 206 that characterize the agent's progress in exploring the environment. The intrinsic rewards 206 generated by the system 300 may be used, for example, by the training system 200 described with reference to FIG.

[0094] The system 300 includes an embedded neural network 302, an external memory 304, and a comparison engine 306, each of which is described in more detail below.

[0095] The embedded neural network 302 is configured to process the current observations 212 to generate an embedding of the current observations 212, referred to as a “controllability representation” 308 (or an “embedded controllability representation”). The controllability representation 308 of the current observations 212 may be represented as an ordered collection of numerical values, e.g., an array of numerical values. The embedded neural network 302 may be implemented as a neural network having multiple layers, where one or more of the layers perform a function defined by weights that are modified during training of the embedded neural network 302. In some cases, particularly when the current observations are in the form of at least one image, one or more of the layers of the embedded neural network, e.g., at least a first layer, may be implemented as a convolutional layer.

[0096] The system 300 may train the embedded neural network 302 to generate controllability representations of observations that characterize aspects of the state of the environment that are controllable by the agent. An aspect of the state of the environment may be referred to as controllable by an agent if it is determined (at least in part) by actions performed by the agent. For example, the position of an object grasped by an actuator of a robotic agent may be controllable by the agent, while the ambient lighting conditions or the movement of other agents in the environment are not controllable by the agent. Exemplary techniques for training the embedded neural network 302 are described in more detail below.

[0097] The external memory 304 stores controllability representations of previous observations that characterize the state of the environment at the previous time step.

[0098] The comparison engine 306 is configured to generate the intrinsic reward 206 by comparing the controllability representation 308 of the current observation 212 with the controllability representations of previous observations stored in the external memory 304. In general, the comparison engine 306 may generate a higher intrinsic reward 206 if the controllability representation 308 of the current observation 212 is less similar to the controllability representations of previous observations stored in the external memory.

[0099] For example, the comparison engine 306 may t of,

number

number

number

number

[0100] Determining the intrinsic reward 206 based on a controllability representation that characterizes controllable aspects of the state of the environment may enable more effective training of the action selection neural network. For example, the state of the environment may vary independently of the actions performed by the agent, e.g., in the case of a real-world environment having variations in lighting and the presence of obstructing objects. In particular, the observations that characterize the current state of the environment may differ substantially from the observations that characterize the previous state of the environment even if the agent has not performed an intervening action during the time step. Therefore, an agent trained to maximize an intrinsic reward determined by directly comparing the observations that characterize the state of the environment may not perform significant exploration of the environment, for example, because the agent may receive a positive intrinsic reward without performing an action. In contrast, the system 300 generates intrinsic rewards that motivate the agent to achieve significant exploration of the controllable aspects of the environment.

[0101] In addition to using the controllability representation 308 of the current observation 212 to generate the intrinsic reward 206 for the current time step, the system 300 may store the controllability representation 308 of the current observation 212 in the external memory 304.

[0102] In some implementations, the external memory 304 may be a "temporary" memory, i.e., the system 300 "resets" the external memory (e.g., by erasing its contents) each time a memory reset criterion is met. For example, the system 300 may determine that the memory reset criterion is met in a current time step if the memory reset criterion was last met a predetermined number N>1 of time steps prior to the current time step, or if the agent completes its task in the current time step. In implementations in which the external memory 304 is a temporary memory, the intrinsic reward 206 generated by the comparison engine 306 may be referred to as a "temporary" intrinsic reward. The temporary intrinsic reward may encourage the agent to continually explore the environment by performing actions that repeatedly transition the state of the environment to each possible state.

[0103] In addition to determining the episodic intrinsic reward, the system 300 may also determine a "non-episodic" intrinsic reward, i.e., an intrinsic reward that depends on the state of the environment for each previous time step, not just those time steps since the last time the episodic memory was reset. The non-episodic intrinsic reward may be, for example, a random network distillation (RND) reward as described in Y. Burda et al., "Exploration by random network distillation," arXiv:1810.12894v1, (2018). The non-episodic intrinsic reward may decrease over time as the agent explores the environment, and does not encourage the agent to repeatedly revisit all possible states of the environment.

[0104] Optionally, the system 300 can generate an intrinsic reward 206 for the current time step based on both the episodic reward and the non-epistatic reward. For example, the system 300 can generate an intrinsic reward R for the time step based on both the episodic reward and the non-epistatic reward. t of,

number

number

number

[0105] Some example techniques for training the embedded neural network 302 are described in more detail below.

[0106] In one example, the system 300 may jointly train the embedded neural network 302 with an action prediction neural network. The action prediction neural network may be configured to receive inputs including respective controllability representations (generated by the embedded neural network) of (i) a first observation characterizing a state of the environment at a first time step and (ii) a second observation characterizing the state of the environment at a next time step. The action prediction neural network may process the inputs to generate predictions for actions performed by the agent that transitioned the state of the environment from the first observation to the second observation. The system 300 may train the embedded neural network 302 and the action prediction neural network to optimize an objective function that measures the error between (i) a predicted action generated by the action prediction neural network and (ii) a "goal" action actually performed by the agent. In particular, the system 300 may back-propagate the gradient of the objective function through the action prediction neural network to the embedded neural network 302 at each of multiple training iterations. The objective function may be, for example, a cross-entropy objective function. Training an embedded neural network in this manner encourages the controllability representation to encode information about the environment that is affected by the agent's actions, i.e., the controllable aspects of the environment.

[0107] In another example, the system 300 can jointly train the embedded neural network 302 with a state predictive neural network. The state predictive neural network can be configured to process inputs including (i) controllability representations of observations (generated by the embedded neural network 302) that characterize the state of the environment at a time step, and (ii) representations of actions performed by the agent at the time step. The state predictive neural network can process the inputs to generate outputs that characterize the predicted state of the environment at the next step, i.e., after the agent has performed the action. The outputs can include, for example, predicted controllability representations that characterize the predicted state of the environment at the next time step. The system 300 can jointly train the embedded neural network 302 and the state predictive neural network to optimize an objective function that measures the error between (i) the predicted controllability representations generated by the state predictive neural network and (ii) a “target” controllability representation that characterizes the actual state of the environment at the next time step. The "goal" controllability representation may be generated by the embedded neural network based on observations that characterize the actual state of the environment at the next time step. In particular, the system 300 may back-propagate the gradient of the objective function through the state prediction neural network to the embedded neural network 302 at each of multiple training iterations. The objective function may be, for example, a squared error objective function. Training the embedded neural network in this manner encourages the controllability representation to encode information about the environment that is affected by the agent's actions, i.e., the controllable aspects of the environment.

[0108] 4 is a flow diagram of an exemplary process 400 for executing a task episode and updating a return calculation scheme selection policy. For convenience, the process 400 is described as being performed by one or more computer systems located in one or more locations. For example, an action selection system, such as the action selection system 100 of FIG. 1, suitably programmed in accordance with this specification, can perform the process 500.

[0109] The system maintains policy data that defines a policy for selecting between a number of different return computation schemes, each of which assigns a different importance to exploring the environment while executing an episode of the task (step 402).

[0110] The system uses the policy to select a return calculation scheme from a number of different return calculation schemes (step 404). For example, the policy may assign a respective reward score to each scheme that represents a current estimate of the reward signal that will be received if the scheme is used to control the agent for an episode. As an example, to select a policy, the system may select the scheme with the highest reward score. As another example, with probability ε, the system may select a random scheme from the set of schemes, and with probability 1-ε, the system may select the scheme with the highest reward score as defined by the policy. As another example, to select a policy, the system may map reward scores over the set of return calculation schemes to a probability distribution and then sample a scheme from the probability distribution.

[0111] The system controls the agent to execute an episode of the task to maximize the return calculated according to the selected return calculation scheme (step 406). That is, during the task episode, the system trains the action selection neural network on data identifying the selected return calculation scheme.

[0112] The system identifies rewards generated as a result of the agent performing the episode of the task, step 408. As a particular example, the system may identify extrinsic rewards, i.e., rewards received at each of the time steps within the task episode that measure progress on the task.

[0113] The system uses the identified rewards to update the policy to select between a number of different return calculation schemes (step 410).

[0114] In general, the system updates the policy using a non-stationary multi-armed bandit algorithm with each arm corresponding to each of the return calculation schemes.

[0115] More specifically, the system can generate a reward signal for the bandit algorithm from the identified rewards and then use the reward signal to update the policy. The reward signal can be a combination of the extrinsic rewards received during the task episode, for example, an undiscounted extrinsic reward that is the undiscounted sum of the received rewards.

[0116] The system can perform the updates using any of a variety of non-stationary multi-armed bandit algorithms.

[0117] As a particular example, the system may calculate, for each scheme, the empirical mean of the reward signals received for episodes within some fixed number of task episodes of the current episode, i.e., within a fixed-length recent planning horizon. The system may then calculate a reward score for each scheme from the empirical means for the schemes. For example, the system may calculate a reward score for a given scheme by adding a confidence bound bonus to the empirical mean for the given scheme. The confidence bound bonus may be determined based on how many times a given scheme has been selected within the recent planning horizon, i.e., schemes that have been selected less times are assigned a larger bonus. As a particular example, the bonus for a given scheme a calculated after the kth task episode may be:

number

[0118] Thus, the system adaptively modifies a policy for selecting between return computation schemes during training of the action selection neural network, resulting in different return computation schemes being preferred (and therefore more likely to be selected) by the policy at different times during training. Because the policy is based on expected reward signals (based on extrinsic rewards) for the different schemes at any given point during training, the system is more likely to select a scheme that is more likely to result in higher extrinsic rewards being collected over a task episode, resulting in higher quality training data being generated for the action selection neural network.

[0119] This specification uses the term "configured" in connection with systems and computer program components. A system of one or more computers to be configured to perform a particular operation or action means that the system has installed thereon software, firmware, hardware, or a combination thereof that, during operation, causes the system to perform the operation or action. A computer program or programs to be configured to perform a particular operation or action means that the program or programs include instructions that, when executed by a data processing device, cause the device to perform the operation or action.

[0120] The subject matter and functional operations described herein, including the structures disclosed herein and their structural equivalents, may be implemented in digital electronic circuitry, tangibly embodied computer software or firmware, computer hardware, or a combination of one or more of them. The subject matter described herein may be implemented as one or more computer programs, i.e., one or more modules of computer program instructions encoded on a tangible, non-transitory storage medium for execution by or to control the operation of a data processing apparatus. The computer storage medium may be a machine-readable storage device, a machine-readable storage substrate, a random or serial access memory device, or a combination of one or more of them. Alternatively or additionally, the program instructions may be encoded on an artificially generated propagated signal, e.g., a mechanically generated electrical, optical, or electromagnetic signal, that is generated to encode information for transmission to a receiver device suitable for execution by a data processing apparatus.

[0121] The term "data processing apparatus" refers to data processing hardware and encompasses any kind of apparatus, device, and machine for processing data, including, by way of example, a programmable processor, computer, or multiple processors or computers. An apparatus may also be or further include special purpose logic circuitry, such as an FPGA (field programmable gate array) or an ASIC (application specific integrated circuit). An apparatus may optionally include, in addition to hardware, code that creates an execution environment for a computer program, such as code constituting a processor firmware, a protocol stack, a database management system, an operating system, or any combination of one or more of these.

[0122] A computer program, which may be referred to or written as a program, software, software application, app, module, software module, script, or code, may be written in any form of programming language, including compiled or interpreted languages, or declarative or procedural languages, and the computer program may be distributed in any form, including as a stand-alone program or as a module, component, subroutine, or other unit suitable for use in a computing environment. A program may correspond to a file in a file system, but this is not necessary. A program may be stored as part of a file that holds other programs or data, for example, in one or more scripts stored in a markup language document, in a single file dedicated to the program, or in multiple cooperating files, for example, files that store one or more modules, subprograms, or portions of code. A computer program may be deployed to run on one computer or multiple computers, located at one site or distributed across multiple sites and interconnected by a data communication network.

[0123] The term "engine" is used broadly herein to refer to a software-based system, subsystem, or process programmed to perform one or more specific functions. Generally, an engine is implemented as one or more software modules or components that are installed in one or more locations on one or more computers. In some cases, one or more computers are dedicated to a particular engine, and in other cases, multiple engines may be installed and run on the same one or more computers.

[0124] The processes and logic flows described herein may be performed by one or more programmable computers executing one or more computer programs to perform functions by operating on input data to generate output. The processes and logic flows may also be performed by special purpose logic circuitry, such as an FPGA or ASIC, or a combination of special purpose logic circuitry and one or more programmed computers.

[0125] A computer suitable for executing a computer program can be based on a general purpose or special purpose microprocessor, or a general purpose and special purpose microprocessor, or any other type of central processing unit. Generally, the central processing unit will receive instructions and data from a read-only memory or a random access memory, or both. The essential elements of a computer are a central processing unit for executing or implementing instructions, and one or more memory devices for storing instructions and data. The central processing unit and memory may be supplemented by or incorporated in special purpose logic circuitry. Generally, a computer also includes one or more mass storage devices, such as magnetic disks, magneto-optical disks, or optical disks, for storing data, or is operatively coupled to receive data from or communicate data to the storage device, or both. However, a computer does not necessarily have to have such devices. Moreover, a computer may be incorporated in another device, such as a mobile phone, a personal digital assistant (PDA), a mobile audio or video player, a game console, a global positioning system (GPS) receiver, or a portable storage device such as a universal serial bus (USB) flash drive, to name a few.

[0126] Computer readable media suitable for storing computer program instructions and data include all forms of non-volatile memory, media and memory devices including, by way of example, semiconductor memory devices such as EPROM, EEPROM and flash memory devices, magnetic disks such as internal hard disks or removable disks, magnetic optical disks, and CD-ROM and DVD-ROM disks.

[0127] To provide for interaction with a user, embodiments of the subject matter described herein may be implemented on a computer having a display device, e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor, for displaying information to the user, and a keyboard and pointing device, e.g., a mouse or trackball, by which the user can provide input to the computer. Other types of devices may be used to provide interaction with the user as well, e.g., feedback provided to the user may be any form of sensory feedback, e.g., visual feedback, auditory feedback, or tactile feedback, and input from the user may be received in any form, including acoustic, speech, or tactile input. In addition, the computer may interact with the user by sending documents to and receiving documents from a device used by the user, e.g., by sending a web page to a web browser on the user's device in response to a request received from the web browser. Similarly, the computer may interact with the user by sending text messages or other forms of messages to a personal device, e.g., a smartphone running a messaging application, and receiving a response message from the user in return.

[0128] A data processing device for implementing machine learning models may also include dedicated hardware accelerator units, for example for handling common computationally intensive parts of machine learning training or production, i.e., inference, workload.

[0129] The machine learning model may be implemented and deployed using a machine learning framework, for example, the TensorFlow framework, the Microsoft Cognitive Toolkit framework, the Apache Singa framework, or the Apache MXNet framework.

[0130] An embodiment of the subject matter described herein may be implemented in a computing system that includes back-end components, e.g., as a data server, or includes middleware components, such as an application server, or includes front-end components, such as a client computer having a graphical user interface, web browser, or app through which a user may interact with an implementation of the subject matter described herein, or in any combination of one or more such back-end, middleware, or front-end components. The components of the system may be interconnected by any form or medium of digital data communication, e.g., a communication network. Examples of communication networks include a local area network (LAN) and a wide area network (WAN), e.g., the Internet.

[0131] A computing system may include clients and servers. Clients and servers are generally remote from each other and typically interact through a communication network. The relationship of client and server arises by virtue of computer programs running on the respective computers and having a client-server relationship to each other. In some embodiments, a server transmits data, e.g., HTML pages, to a user device, e.g., to display data to and receive user input from a user interacting with the device acting as a client. Data generated at the user device, e.g., a result of a user interaction, may be received from the device at the server.

[0132] Although this specification contains many specific implementation details, these should not be construed as limiting the scope of the invention or what is claimed, but as describing features specific to a particular embodiment of a particular invention. Some features described in the context of a separate embodiment may also be implemented in combination in a single embodiment. Conversely, various features described in the context of a single embodiment may also be implemented in multiple embodiments individually or in any suitable subcombination. Moreover, although features are described above as working in a particular combination and are initially claimed as such, in some cases one or more features from a claimed combination may be deleted from the combination, or the claimed combination may be transferred to a subcombination or a variation of the subcombination.

[0133] Similarly, although operations are shown in the figures and recited in the claims in a particular order, this should not be understood as requiring that such operations be performed in the particular order shown, or in sequential order, or that all of the illustrated operations be performed, to achieve desirable results. In some situations, multitasking and parallel processing may be advantageous. Moreover, the separation of various system modules and components in the embodiments described above should not be understood as requiring such separation in all embodiments, and it should be understood that the program components and systems described may generally be integrated together in a single software product or packaged in multiple software products.

[0134] Particular embodiments of the subject matter have been described. Other embodiments are within the scope of the following claims. For example, the actions recited in the claims can be performed in a different order and still achieve desirable results. As an example, the processes depicted in the accompanying figures do not necessarily require the particular order depicted, or sequential order, to achieve desirable results. In some cases, multitasking and parallel processing may be advantageous. [Explanation of symbols]

[0135] 100 Action Selection System 102 Action Selection Neural Network 104 Agent 106 Environment 108 Action 110 observations 112 Return Calculation Method 114 Action Score 120 Policy Data 130 Extrinsic rewards 132 Intrinsic Rewards 200 Training System 202 Total Rewards 204 Extrinsic rewards 206 Intrinsic rewards, exploration rewards 208 Training Engine 210 Return Calculation Method 212 observations 300 Intrinsic Reward System 302 Embedded Neural Networks 304 External Memory 306 Comparison Engine 308 Controllability Expression

Claims

1. 1. A method implemented by one or more computers for controlling an agent that interacts with an environment to perform episodes of a task, comprising: maintaining data defining a policy for selecting between a plurality of different return calculation schemes, each return calculation scheme assigning a different importance to exploring the environment while performing the episode of the task, the policy assigning each of the return calculation schemes a respective reward score; selecting a return calculation scheme from the plurality of different return calculation schemes using the policy; controlling the agent to perform the episodes of the task to maximize a return calculated according to the selected return calculation scheme; identifying a reward generated as a result of the agent performing the episode of the task; updating the policy using the identified reward to select between a plurality of different return calculation schemes; A method comprising:

2. 2. The method of claim 1, wherein the plurality of different return calculation formulas each define at least a respective discount factor used in combining fees to generate a return.

3. 3. The method of claim 1 or 2, wherein the plurality of different return calculation schemes each define at least one respective intrinsic reward scaling factor that defines the importance of intrinsic rewards relative to extrinsic rewards received from the environment when generating a return.

4. controlling the agent to perform the episode of the task to maximize a return calculated according to the selected return calculation scheme, comprising: receiving observations characterizing a current state of the environment; processing the observations and data defining the selected return calculation scheme using one or more action selection neural networks to generate an action selection output; selecting an action to be performed by the agent using the action selection output; The method according to claim 1 , further comprising the step of iteratively performing:

5. 5. The method of claim 4, wherein the environment is a real-world environment, each observation is an output of at least one sensor configured to sense the environment, and the agent is a mechanical agent that interacts with the environment.

6. 1. A method implemented by one or more computers for controlling an agent that interacts with an environment to perform episodes of a task, comprising: maintaining data defining a policy for selecting between a plurality of different return calculation schemes, each return calculation scheme assigning a different importance to exploring the environment while executing the episodes of the task; selecting a return calculation scheme from the plurality of different return calculation schemes using the policy; controlling the agent to perform the episodes of the task to maximize a return calculated according to the selected return calculation scheme; receiving observations characterizing a current state of the environment; processing the observations and data defining the selected return calculation scheme using one or more action selection neural networks to generate an action selection output; selecting an action to be performed by the agent using the action selection output; a controlling step including repeatedly executing identifying a reward generated as a result of the agent performing the episode of the task; updating the policy using the identified reward to select between a plurality of different return calculation schemes; Including, the one or more action selection neural networks an intrinsic reward action selection neural network that estimates an intrinsic return calculated solely from intrinsic rewards generated by an intrinsic reward system based on observations received during interaction with the environment; and an extrinsic reward action selection neural network that estimates an extrinsic return calculated solely from extrinsic rewards received from the environment as a result of interaction with the environment.

7. processing the observations and data defining the selected return calculation scheme using one or more action selection neural networks to generate an action selection output, the step including, for each action in the set of actions: processing the observations, the actions, and the data defining the selected return calculation scheme using the intrinsic reward action selection neural network to generate an estimated intrinsic return that will be received if the agent performs the action in response to the observations; processing the observations, the actions, and the data defining the selected return calculation scheme using the extrinsic reward action selection neural network to generate an estimated extrinsic return that will be received if the agent performs the action in response to the observations; and determining a final return estimate from the estimated intrinsic reward and the estimated extrinsic reward.

8. selecting an action to be performed by the agent using the action selection output, selecting, with probability 1-ε, an action corresponding to a highest final return estimate among the final return estimates determined for each action in the set of actions, and selecting a random action from the set of actions with probability ε.

9. 9. The method of claim 6, wherein the two action selection neural networks have the same architecture but different parameter values.

10. generating training data from performance of episodes of the task; and training the one or more action selection neural networks on the training data through reinforcement learning.

11. training the one or more action selection neural networks on the training data, training the intrinsic reward action selection neural network using only intrinsic rewards generated as a result of the performance of episodes of the task; and training the extrinsic reward action selection neural network using only extrinsic rewards received during the performance of episodes of the task.

12. 1. A method implemented by one or more computers for controlling an agent that interacts with an environment to perform episodes of a task, comprising: maintaining data defining a policy for selecting between a plurality of different return calculation schemes, each return calculation scheme assigning a different importance to exploring the environment while executing the episodes of the task; selecting a return calculation scheme from the plurality of different return calculation schemes using the policy; controlling the agent to perform the episodes of the task to maximize a return calculated according to the selected return calculation scheme; identifying a reward generated as a result of the agent performing the episode of the task; updating the policy using the identified rewards to select between a plurality of different return calculation schemes, the policy being updated using a non-stationary multi-armed bandit algorithm having a respective arm corresponding to each of the return calculation schemes; A method comprising:

13. using the identified reward to update the policy to select among a plurality of different return calculation schemes; determining an undiscounted extrinsic return from extrinsic rewards received during performance of an episode of the task; and updating the policy by using the undiscounted extrinsic reward as a reward signal for the non-stationary multi-armed bandit algorithm.

14. 14. A system including one or more computers and one or more storage devices storing instructions that, when executed by the one or more computers, cause the one or more computers to perform operations of the method of any one of claims 1 to 13.

15. One or more non-transitory computer storage media storing instructions that, when executed by one or more computers, cause the one or more computers to perform operations of the method of any one of claims 1 to 13.

Citation Information

Patent Citations

  • Environmental prediction using reinforcement learning

    JP2019537136A