TLDR: GNOME (Generating Novelty in Open-world Multi-agent Environments) is a simulation platform designed to test how multi-agent AI systems adapt to unexpected changes, or “novelty,” in strategic board games like Monopoly. Developed under the DARPA SAIL-ON program, GNOME allows researchers to inject various types of novelties (e.g., changes in game rules, object attributes, or environment layout) into the game in real-time, evaluating agents’ ability to detect and react without retraining. This platform aims to advance AI research towards more general intelligence capable of operating in dynamic, open-world environments.
Artificial intelligence (AI) has made incredible strides in mastering complex games, from Chess and Go to Starcraft. However, a significant challenge remains: how well do these AI systems perform when faced with unexpected changes, or ‘novelty,’ in their environment? A new research paper introduces GNOME (Generating Novelty in Open-world Multi-agent Environments), an experimental platform designed to tackle this very question.
Traditionally, AI game-playing agents are developed with the assumption that the game’s rules and environment remain static. Their variance comes only from stochastic elements like dice rolls or other agents’ decisions. Yet, real-world environments are dynamic and unpredictable. An AI that truly understands a domain should be able to detect and adapt to novelties without needing to be retrained from scratch.
GNOME, developed by Mayank Kejriwal and Shilpa Thomas from the Information Sciences Institute at the University of Southern California, aims to provide a robust simulator for systematic experiments in this new area of AI research. It separates the AI agent’s development from the simulator, allowing for the injection of unanticipated novelty. This platform is funded under the DARPA Science of Artificial Intelligence and Learning for Open-World Novelty (SAIL-ON) program, which seeks to uncover the scientific principles and algorithms for training agents that can act effectively in novel, open-world situations.
Monopoly as the Testbed
While GNOME is designed to support various multi-agent strategic board games, the classic game of Monopoly is used to illustrate its capabilities in this paper. Monopoly, with its vast decision space and mix of relevant and irrelevant elements, serves as an ideal environment to test AI robustness against change. The game involves four players, dice rolls, property acquisition, trading, and elements of chance and community chest.
Understanding Novelty
The paper defines novelty as situations that violate an agent’s implicit or explicit assumptions about the external world. To manage this broad concept, GNOME categorizes novelty into several types:
-
Attribute Novelty: Changes to an object’s attributes. For example, a property’s color might change, altering its monopoly group and strategic value. If ‘Boardwalk’ turns lime-green and is now a unique color, it can be improved without acquiring other properties, a significant shift for an adaptive agent.
-
Class Novelty: Introduction of previously unseen objects or entities. An example is adding a third die to the game, which changes movement probabilities and impacts the value of different strategies, like focusing on railroads over properties.
-
Representation Novelty: Changes in how entities and features are specified, like transforming dimensions or coordinate systems. For Monopoly, this could mean scrambling the board layout or extending the number of slots for certain locations, such as tax spaces, leading to increased financial risk.
The GNOME Demonstration at NeurIPS 2020
At NeurIPS 2020, GNOME was demonstrated via a web-based graphical user interface (GUI). The workflow allowed participants to:
-
Select Agent Combinations: Users could choose from a library of pre-programmed agents, including a ‘simple’ agent, two ‘heuristic’ agents (H1 and H2) with progressively sophisticated hard-coded rules, and a ‘hybrid ML’ agent trained using reinforcement learning for some decisions.
-
Inject Novelty: Participants could select from a curated set of novelties, such as increasing the number of dice, changing property colors to create new monopoly dynamics, or extending the number of slots for specific locations like tax spaces.
-
Visualize Gameplay: The game’s evolution with the injected novelty could be observed in a 2D gameboard GUI, showing player positions, cash, and property ownership. This allowed for direct comparison of how different agents reacted to the same novelty.
Evaluating Novelty Adaptation
GNOME is publicly available and used in the DARPA SAIL-ON program to evaluate externally developed agents. The experimental protocol involves tournaments where agents play a sequence of games. A ‘pre-novelty’ phase establishes baseline performance, followed by a ‘post-novelty’ phase where novelty is injected at the start of each game. The primary performance metric is the ‘Win Ratio,’ tracking the fraction of games won by the agent. Agents are also required to emit a binary signal when they detect a novelty, allowing for evaluation of detection speed.
The goal is for agents to adapt in real-time, often within just a few games, without going ‘offline’ for retraining. The novelties used in evaluations are designed to be unanticipated, ensuring that only agents with a deep, general understanding of the domain can robustly adapt.
Also Read:
- Building a Collaborative Space for AI in Network Troubleshooting
- TALENTS: An AI Framework for Adaptive Zero-Shot Human-Agent Collaboration
The Path Forward
Future work for GNOME includes exploring more advanced types of novelty, such as ‘interaction novelties’ (e.g., allowing agents to collude), ‘environment novelties’ (e.g., playing a different regional version of Monopoly), and ‘game rule novelties’ that fundamentally alter the game. The long-term vision is to extend this framework to other multi-agent games like Poker, establishing a general experimental platform for AI research focused on novelty. This research, detailed further in the paper available at arXiv:2507.03802, is crucial for developing AI systems that can truly thrive in the unpredictable complexities of the real world.
