An Open-World Novelty Generator for Authoring Reinforcement Learning Environment of Standardized Toolkits

Abstract

Current research in Reinforcement Learning (RL) is based on closed-world learning environment where the environment remains fixed and unchanged throughout the agent’s training and application session. The fixed environment may be prone to failure when the agents incorporate under novel unseen situations. To overcome the drawback of the existing closed-world model, an Open-world learning model is required which can classify the novelty occurring in an environment in a hierarchical manner. The proposed control suite with open world novelty generator is an attempt to augment the machine learning environment for authoring the novelty in actors, interactions, and environment of standardized Reinforcement learning toolkits such as UnityML, OpenAI Gym, and DeepMind Control Suite in real-time. Such a tool will provide an opportunity to the RL researchers to simulate the Open-world learning model and test their algorithms within the standardized closed-world learning environments of the standardized RL toolkits.