Multi-task neural network systems with task-specific policies and a shared policy

A method of training an adaptive multitask computer system, the computer system comprising a set of trainable workers and a shared module, the method comprising:

• coordinated training of said trainable workers and said shared module on a plurality of different tasks, such that each worker learns to perform a corresponding one of the tasks according to a respective task policy and said shared policy network learns a multitask policy representing common behavior for the tasks;
• and
• wherein said coordinated training is performed by optimizing an objective function, the objective function comprising, for each task,
• a reward term indicative of an expected reward earned by a worker in performing the task;
• and
• at least a first entropy term measuring a difference between a distribution of said task policy and a distribution of said multitask policy to regularize the distribution of the task policy towards the distribution of the multitask policy.

A multitask neural network system to process input data relating to a plurality of learned tasks to be performed and to provide output data to control performance of the tasks, the system comprising:

• an input to receive said input data;
• a set of workers, coupled to said input, each configured to perform an associated one of the plurality of learned tasks by outputting action data according to a respective learned task policy, and each task being associated with a control module;
• and
• a shared policy network, coupled to one or both of said input and said set of workers, wherein said shared policy network defines a learned multitask policy representing common behavior for the tasks;
• and
• wherein multitask neural network system is configured, for each task, to combine an output of the associated control module with said learned multitask policy to define the respective task policy.