A small program that automatically generates simple meta-reinforcement learning tasks. The resulting space of accessible meta-tasks is very large and includes bandit tasks, the Harlow task, the two-step task, T-mazes, and other known meta-tasks from the literature.
A detailed description is available at https://arxiv.org/abs/2302.05583.
Here is a description of the Daw two-step task in the parametrization used by this program, as explained in the preprint:
It is recommended to first consult simple.py, which is simplified as much as
possible for illustrative purposes. Running the script generates and prints out
the (textual) specification for one meta-task. Conventions are included in the printout.
The code actually used for task generation is tasks.py. This code contains
various tricks and workarounds to bias the generative process towards
(hopefully) more interesting or interpretable meta-tasks.
The file ExampleTasks.txt contains the generated specifications for the three
randomly generated meta-tasks described in the preprint.
Work in progress.