Generating Novelty in Open-world Multi-agent Environments


GNOME is a computational platform for generating, studying and analyzing novelty in the context of open-world Artificial Intelligence (AI). Open-world intuitively means that the world is not fully specified, known or parameterized in advance. The real world, as seen from the human perspective, is the best example. In the real world, human beings are constantly having to adapt to novelty, both in the long and short term. In contrast, AI systems are far less versatile, despite having achieved near human-level performance in specific, highly scoped problem areas such as face recognition. As a platform, GNOME allows researchers to simulate and 'play with' novelty as a first-class citizen. GNOME also serves as a testbed for evaluating novelty-adaptive AI agents in strategic gameplaying environments.

GNOME is funded under the DARPA SAIL-ON (Science of Artificial Intelligence and Learning for Open-world Novelty) program. While current AI systems excel at tasks defined by 'rigid' rules, they are not very good at adapting to changing conditions commonly faced by people in the real world: from driving in fluctuating weather, walking on uneven terrain, playing games with slightly different rules (e.g., chess under a time constraint) and, in general, modifying the internal model of the world when an assumption turns out to be faulty.

Currently, GNOME has been implemented for the Monopoly board game, and includes a novelty simulator. The package and simulator is publicly available on GitHub. GNOME is a collaboration with researchers from Purdue University.






This work was funded by the Defense Advanced Research Projects Agency with award W911NF2020003.