Deepmind Technologies Limited

Hado Philip Van Hasselt

David Silver

Tom Schaul

Satinder Singh Baveja

Arthur Clement Guez

Matteo Hessel

Zhongwen Xu

Arthur Clément Guez

Junhyuk Oh

Daniel J Mankowitz

Iurii Kemaev

Tom Ben Zion Zahavy

Vivek Veeriah Jeya Veeraiah

A method of training an action selection neural network for controlling an agent interacting with an environment to perform different tasks is described. The method includes obtaining a first trajectory of transitions generated while the agent was performing an episode of the first task from multiple tasks; and training the action selection neural network on the first trajectory to adjust the control policies for the multiple tasks. The training includes, for each transition in the first trajectory: generating respective policy outputs for the initial observation in the transition for each task in a subset of tasks that includes the first task and one other task; generating respective target policy outputs for each task using the reward in the transition, and determining an update to the current parameter values based on, for each task, a gradient of a loss between the policy output and the target policy output for the task.

Continual reinforcement learning with a multi-task agent

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for prediction of an outcome related to an environment. In one aspect, a system comprises a state representation neural network that is configured to: receive an observation characterizing a state of an environment being interacted with by an agent and process the observation to generate an internal state representation of the environment state; a prediction neural network that is configured to receive a current internal state representation of a current environment state and process the current internal state representation to generate a predicted subsequent state representation of a subsequent state of the environment and a predicted reward for the subsequent state; and a value prediction neural network that is configured to receive a current internal state representation of a current environment state and process the current internal state representation to generate a value prediction.

Environment prediction using reinforcement learning

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for training a Q network used to select actions to be performed by an agent interacting with an environment. One of the methods includes obtaining a plurality of experience tuples and training the Q network on each of the experience tuples using the Q network and a target Q network that is identical to the Q network but with the current values of the parameters of the target Q network being different from the current values of the parameters of the Q network.

Training reinforcement learning neural networks

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for reinforcement learning. The embodiments described herein apply meta-learning (and in particular, meta-gradient reinforcement learning) to learn an optimum return function G so that the training of the system is improved. This provides a more effective and efficient means of training a reinforcement learning system as the system is able to converge on an optimum set of one or more policy parameters &#x3b8; more quickly by training the return function G as it goes. In particular, the return function G is made dependent on the one or more policy parameters &#x3b8; and a meta-objective function J' is used that is differentiated with respect to the one or more return parameters &#x3b7; to improve the training of the return function G.

Meta-gradient updates for training return functions for reinforcement learning systems

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for training a neural network using normalized target outputs. One of the methods includes updating current values of the normalization parameters to account for the target output for the training item; determining a normalized target output for the training item by normalizing the target output for the training item in accordance with the updated normalization parameter values; processing the training item using the neural network to generate a normalized output for the training item in accordance with current values of main parameters of the neural network; determining an error for the training item using the normalized target output and the normalized output; and using the error to adjust the current values of the main parameters of the neural network.

Growing community of inventors

London, United Kingdom

Hado Philip Van Hasselt