My project aims to explore how artificial intrinsic motivation can encourage agent exploration in large action spaces with sparse rewards. In order for an autonomous agent to successfully navigate a large action space and achieve a wide range of user-specified tasks at test time, it must first be able to create a large repertoire of general-purpose skills. That said, navigating a large action space with sparse rewards can be extremely challenging, intrinsic reward signals have been used to combat this issue and encourage agent exploration. Furthermore, raw sensory inputs such as images can be used to improve the agent’s ability to generalize its acquired skills to different domains. My project aims to develop algorithms that encourage agent exploration by combining standard reinforcement learning, computer vision, and outlier detection methods.
My proposed research chooses to learn a visual representation that serves three different purposes: providing a structured encoded representation of the raw sensory inputs, identifying and sampling outlier values for self-supervised goal labeling, and computing a reward signal as the agent nears the novel goal. The overarching intention is that the proposed research is efficient enough to encourage the acquisition of general-purpose skills and extend past simulations to learn policies that can operate on raw image sensory inputs and goals for real-world robotic systems.
Artificial Intelligence and Machine Learning with a focus in Reinforcement Learning, Unsupervised Learning and Cognitive Development of Robots. Above all, the ethical and transparent implementation of these subject matters.
BSc Economics and Women & Gender Studies, Rutgers University
MSc Data Science, University of Bath
I have three years of experience working as a software and management consultant, and one year of experience as a data analyst within the fintech and blockchain domain.
Prof Özgür Şimşek
Dr Gosia Gocłowska
