notesum.ai

Published at October 18

Interpretable end-to-end Neurosymbolic Reinforcement Learning agents

cs.CL

cs.AI

Released Date: October 18, 2024

Authors: Nils Grandien¹, Quentin Delfosse², Kristian Kersting³

Aff.: ¹Computer Science Department, TU Darmstadt, Germany; ²Computer Science Department, TU Darmstadt, Germany; National Research Center for Applied Cybersecurity Darmstadt, Germany; ³Computer Science Department, TU Darmstadt, Germany; Hessian Center for Artificial Intelligence (hessian.AI), Darmstadt, Germany; Centre for Cognitive Science, TU Darmstadt, Germany; German Research Center for Artificial Intelligence (DFKI), Darmstadt, Germany

Arxiv: https://arxiv.org/abs/2410.14371v1

Pong	unpruned		pruned
	2 layer	1 layer	2 layer	1 layer
neural ground truth objects	$17.4\pm\scriptsize{1.6}$	$17.0\pm\scriptsize{2.3}$	$19.0\pm\scriptsize{1.9}$	$14.6\pm\scriptsize{1.0}$
rule set ground truth objects	$4.8\pm\scriptsize{5.7}$	$-15.8\pm\scriptsize{3.1}$	$15.0\pm\scriptsize{3.0}$	$10.6\pm\scriptsize{2.4}$
rule set SPACE+MOC objects	$-7.6\pm\scriptsize{9.2}$	$-3.2\pm\scriptsize{6.3}$	$14.4\pm\scriptsize{2.6}$	$-3.8\pm\scriptsize{3.4}$
neural SPACE+MOC objects	$16.8\pm\scriptsize{1.5}$	$5.4\pm\scriptsize{3.1}$	$16.8\pm\scriptsize{2.3}$	$-1.0\pm\scriptsize{6.6}$
image data	16.4
random	-20.7
human	9.3