Exploring Behavior via Neural Network Activations in Deep Reinforcement Learning Agents

conference-abstract
A conference abstract on linking neural activation patterns to behavioral components in deep reinforcement learning agents.
Authors

Melanie Lu

Richard Meyes

Andres-Felipe Posada-Moreno

Tobias Meisen

Published

July 26, 2021

Resources

Conference abstract (p. 73) | Conference program

Publication history

Listed in the ICAI 2021 program and published as a conference abstract on page 73 of the CSCE 2021 Book of Abstracts. The linked resources provide the abstract and program; a publicly accessible full manuscript has not been located as of 8 October 2026.

Summary

The conference abstract describes an analysis of deep reinforcement learning agents in continuous motor-control environments. It relates the activation of individual neurons and neuron groups to recurring movements during task execution, examining these associations across network topologies, initializations and control domains. The work draws inspiration from studies of functional neuron populations in the motor cortex.

Citation

@inproceedings{lu2021exploringbehavior,
 author = {Lu, Melanie and Meyes, Richard and Posada-Moreno, Andres-Felipe and Meisen, Tobias},
 title = {Exploring Behavior via Neural Network Activations in Deep Reinforcement Learning Agents},
 booktitle = {CSCE 2021 Book of Abstracts},
 year = {2021},
 pages = {73},
 publisher = {American Council on Science and Education},
 isbn = {1-60132-514-2},
 note = {Conference abstract; ICAI 2021},
 url = {https://www.american-cse.org/static/CSCE21%20book%20abstracts.pdf#page=73}
}