Alberto Camacho and Rodrigo Toro Icarte with their PhD supervisor Professor Sheila McIlraith at ICAPS 2023 in Prague. (Photo: Supplied)
Two recent PhD graduates of the Department of Computer Science, both supervised by Professor Sheila McIlraith, were honoured for their outstanding doctoral theses at the 2023 International Conference on Automated Planning and Scheduling (ICAPS).
Alberto Camacho (PhD 2022) is one of two winners of the ICAPS Best Dissertation Award 2023 for his dissertation, “Automata-Theoretic Synthesis of Plans and Reactive Strategies.”
Camacho is currently an AI Research Scientist at X, a research and development subsidiary of Alphabet.
The citation associated with the award reads:
“The dissertation investigates the connections between automated planning in fully observable nondeterministic domains (FOND) with temporally extended goals and temporal synthesis with respect to logical specifications expressed in linear temporal logics on finite and infinite traces. It employs concepts from Formal Methods along with techniques and tools from AI planning. Automated planning and temporal synthesis have been two closely related research areas for decades. Historically, temporal synthesis has been pursued by Formal-Methods researchers, while automated planning has been pursued by the AI community. The dissertation offers significant contributions of both foundational and practical nature, with long-term impacts on planning and temporal synthesis. It represents a substantial advance in comprehending the connections between these two fields.”
Rodrigo Toro Icarte (MSc 2018, PhD 2022) received an honourable mention for his dissertation “Reward Machines.”
Toro Icarte is currently an assistant professor in the Department of Computer Science at Pontificia Universidad Católica de Chile.
The citation associated with the honour reads:
“This dissertation introduces the notion of Reward Machines (RM), an automata structure used to represent the reward function, which can serve as an effective tool to address sample efficiency and partial observability, two challenges that reinforcement learning faces. To that end, the thesis proposes a number of different approaches to policy learning in both tabular and deep learning settings and shows, through experiments, that reinforcement learning with RMs can find high quality policies with significant decrease in sample requirements. The thesis also proposes to learn the RM, using it as a form of external memory that is shown effective under partial observability. The dissertation is highly creative, novel, and elegant, both theoretically and experimentally. While the thesis makes strong contributions to reinforcement learning, it also demonstrates the value of using AI Planning techniques to the study of reinforcement learning. As such the dissertation has significant value in that it serves as an important connection between symbolic techniques and sequential decision making in machine learning.”
Camacho and Toro Icarte were presented with their awards at ICAPS 2023 in Prague on July 12.