Michael Behr, Martin Heidenreich, Gert Schmidt und Hans-Alexander Graf von Schwerin (unter Mitarbeit

che Technikgestaltung, Materialien und Berichte, Bd. 18, Opladen: Westdeutscher Verlag, 1991, ISBN 3-531-12252-5, 214 S., DM 32,- review …

Section 6.3, we take the shortest available samples. The second measure prevents the extraction from too readily terminating on small DFAs. Our method for it is presented in Section 7.3.1. 7.3.1. AGGRESSIVE [...] 1.2 (()i)ma 32.6 ((())) 2.1 (((()))) 3.1 ((((())))) 3.8 (((((()))))) 4.4 ((((((())))))) 6.6 (((((((()))))))) 9.2 ((((((((v()))))))) 10.7 ((((((((a()z))))))))) 8.3 and possibly return them at every equivalence [...] Train Set Accuracy Max Nest. Depth Network Abstr RS Abstr RS GRU 99.98 87.12 8 2 LSTM 99.98 94.19 8 3 Table 3. Counterexamples generated during extraction of automata from a GRU network trained on BP. Refinement …

1Label = 0 ... A1 ... ... ............ ... ... ... ... ...... ...A1 A1 A1 A6 A7 A5A3 A3 A5 ...A9 A4A3 A5 A7A6 A9 A4A3 A5 A7A6 Fig. 2. Figure showing how the speculative decision tree is build by a peer [...] the induced tree of depth 5 is around 3% more accurate than the tree of depth 3, we have used trees of depth 3 in all as a baseline. This is because a tree of depth 3 has far lower communication overhead [...] is likely due to over-fitting; since for adepth of 3 the average number of tuples per leaf is 56 compared to only 3 tuples per leaf for depth of 7. 3) The PeDiT algorithm is guaranteed to converge to the …