/WSU-NG

Novelty Generator for the evaluation of novelty-aware AI agents

Primary LanguagePythonMIT LicenseMIT

WSU-NG

Novelty Generator for the evaluation of novelty-aware AI agents

Installation

Quick Start

To build the sample agent:

docker-compose -f evaluator-[cartpole, vizdoom].yml build

To run the sample agent:

docker-compose -f evaluator-[cartpole, vizdoom].yml up

To stop the run midway press CTRL + C.

Confirm processes are terminated:

docker-compose -f evaluator-[cartpole, vizdoom].yml down

Setting parameters

  • To set desired experimental parameters, modify the [sail-on] parameters in TA1.config.

  • Additional parameter variations will be made available shortly.

Adding a custom agent

Dockerized Agents

Custom agents can be easily added through the use of docker. If your agent is already packaged then all you need to do is modify evaluator-[cartpole, vizdoom].yml.

In evaluator-[cartpole, vizdoom].yml under sample-agent::

  • Set build:
    • context: <dir> where <dir> is your dockerfile location.
    • dockerfile: <filename> where <filename> is your dockerfile's name.
  • Set command: <cmd> where <cmd> is your docker entry point command.

For more information on agent integration see the agent README.md.README.md

Python agents

Python based agents can be easily added in. A base random action agent random_action.py is provided as an example starting point. Additional functionality can be used by directly integrating with the instance handler sample-cartpole.py or sample-vizdoom.py.

Notes

  • Both the Cartpole3D and ViZDoom novelty generators startup when calling the evaluation compose file.
  • By default, the base parameters are set to:
    • novelty levels = [200, 101, 102, 103, 104, 105, 106, 107, 108]
    • difficulty = [easy, medium, hard]
    • novelty_visibility = [0, 1]
    • hint_level = [-1, 0, 1, 2, 3]
  • For quicker experiments, reduce the number of the base parameters.
  • When running the system you will see several lines of run() Connection was closed, reconnecting.... This is normal as the subcomponents finish building out of order.

Acknowledgements

This work was supported by the DARPA SAIL-ON program under cooperative agreement number W911NF-20-2-0004.