Vivek Myers

EECS Department, University of California, Berkeley

Technical Report No. UCB/EECS-2025-2

January 4, 2025

http://www2.eecs.berkeley.edu/Pubs/TechRpts/2025/EECS-2025-2.pdf

Intelligent agents must learn effective representations of the world in order to accomplish different objectives. This thesis focuses on the following question: how should intelligent agents represent the world in order to reach their goals? General goal-reaching abilities requires understanding the temporal structure relating states and future observations or tasks in the world. We explore algorithms for learning structured representations of both the state of the world and the task to enable broad generalization capabilities. Topics discussed include how these representations can be made compatible with language and other forms of task abstraction, how state and goal representations can be made consistent with the temporal structure of decision-making problems to enable compositional and long-horizon decision-making, and how representation structure can be leveraged to reason about intrinsic motivation objectives like empowerment and surprise. The empirical results show that these algorithms can effectively learn representations for decision-making settings such as robotic manipulation, assistance, and locomotion.

Advisors: Anca Dragan and Sergey Levine

---

BibTeX citation:

@mastersthesis{Myers:EECS-2025-2,
    Author= {Myers, Vivek},
    Title= {Structured Representations for Goal-Directed Decision Making},
    School= {EECS Department, University of California, Berkeley},
    Year= {2025},
    Month= {Jan},
    Url= {http://www2.eecs.berkeley.edu/Pubs/TechRpts/2025/EECS-2025-2.html},
    Number= {UCB/EECS-2025-2},
    Abstract= {Intelligent agents must learn effective representations of the world in
order to accomplish different objectives. This thesis focuses on the
following question: how should intelligent agents represent the world in
order to reach their goals? General goal-reaching abilities requires
understanding the temporal structure relating states and future
observations or tasks in the world. We explore algorithms for learning
structured representations of both the state of the world and the task
to enable broad generalization capabilities. Topics discussed include
how these representations can be made compatible with language and other
forms of task abstraction, how state and goal representations can be
made consistent with the temporal structure of decision-making problems
to enable compositional and long-horizon decision-making, and how
representation structure can be leveraged to reason about intrinsic
motivation objectives like empowerment and surprise. The empirical
results show that these algorithms can effectively learn representations
for decision-making settings such as robotic manipulation, assistance,
and locomotion.},
}

EndNote citation:

%0 Thesis
%A Myers, Vivek 
%T Structured Representations for Goal-Directed Decision Making
%I EECS Department, University of California, Berkeley
%D 2025
%8 January 4
%@ UCB/EECS-2025-2
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/2025/EECS-2025-2.html
%F Myers:EECS-2025-2