notesum.ai

Published at November 11

'Explaining RL Decisions with Trajectories': A Reproducibility Study

cs.AI

Released Date: November 11, 2024

Authors: Karim Abdel Sadek¹, Matteo Nulli¹, Joan Velja¹, Jort Vincenti¹

Aff.: ¹University of Amsterdam

Arxiv: http://arxiv.org/abs/2411.07200v1

Refer to caption

$\pi$	$\mathbb{E}[V(s_{0})]$	$\mathbb{E}[\|\Delta Q_{\pi_{\text{orig}}}(s)\|]$	$\mathbb{E}[1(\pi_{\text{orig}}(s)\neq\pi_{j}(s))]$	$W_{\text{dist}}(\bar{d},\bar{d}_{j})$	$\mathbb{P}(c_{\text{final}}=c_{j})$
orig	0.307 $\pm$ 2e-04	-	-	-	-
0	0.305 $\pm$ 9e-12	0.011 $\pm$ 0.001	0.041 $\pm$ 0.047	0.241 $\pm$ 0.227	0.000 $\pm$ 0.000
1	0.303 $\pm$ 2e-03	0.007 $\pm$ 0.011	0.041 $\pm$ 0.042	0.371 $\pm$ 0.348	0.025 $\pm$ 0.050
2	0.304 $\pm$ 0e+00	0.002 $\pm$ 0.001	0.122 $\pm$ 0.049	0.924 $\pm$ 0.150	0.0000 $\pm$ 0.000
3	0.304 $\pm$ 0e+00	0.022 $\pm$ 0.001	0.0000 $\pm$ 0.049	0.111 $\pm$ 0.094	0.0000 $\pm$ 0.000
4	0.305 $\pm$ 1e-11	0.041 $\pm$ 0.004	0.122 $\pm$ 0.049	0.142 $\pm$ 0.096	0.0000 $\pm$ 0.000
5	0.300 $\pm$ 3e-03	0.029 $\pm$ 0.005	0.041 $\pm$ 0.055	0.036 $\pm$ 0.017	0.175 $\pm$ 0.231
6	0.287 $\pm$ 5e-03	0.061 $\pm$ 0.015	0.163 $\pm$ 0.049	0.040 $\pm$ 0.051	0.500 $\pm$ 0.237
7	0.301 $\pm$ 7e-03	0.030 $\pm$ 0.016	0.020 $\pm$ 0.059	0.099 $\pm$ 0.178	0.150 $\pm$ 0.183
8	0.305 $\pm$ 6e-13	0.008 $\pm$ 0.010	0.021 $\pm$ 0.049	0.022 $\pm$ 0.026	0.100 $\pm$ 0.200
9	0.304 $\pm$ 1e-11	0.017 $\pm$ 0.011	0.143 $\pm$ 0.043	0.061 $\pm$ 0.095	0.05 $\pm$ 0.061