Table 2
Performance of 11 different 3-layer CNNs.
| CNN design | |||||
|---|---|---|---|---|---|
| Filters | Past frames (k & m) | Inf. speed (VLT/ELT) | Tr. time/episode (VLT) | Strehl/reward (VLT 0-mag) | |
| CNN 1 | 32 | 10 | 0.29/0.35 ms | 1.4s | 95.61/−4101 |
| CNN 2 | 32 | 15 | 0.30/0.37 ms | 1.5/7 (ELT) s | 95.69/−3340 |
| CNN 3 | 32 | 20 | 0.30/0.40 ms | 1.6 s | 95.74/−3029 |
| CNN 4 | 32 | 25 | 0.30/0.43 ms | 1.8 s | 95.75/−2934 |
| CNN 5 | 64 | 10 | 0.30/0.67 ms | 2.0s | 95.60/−4002 |
| CNN 8 | 64 | 15 | 0.31/0.70 ms | 2.2s | 95.75/−3253 |
| CNN 7 | 64 | 20 | 0.31/0.74 ms | 2.5 s | 95.75/−3052 |
| CNN 8 | 64 | 25 | 0.32/0.79 ms | 2.5 s | 95.76/−2845 |
| CNN 9 | 128 | 10 | 0.36/1.52 ms | 3.7 s | 95.65/−3656 |
| CNN 10 | 128 | 15 | 0.37/1.58 ms | 3.8 s | 95.71/−2943 |
| CNN 11 | 128 | 20 | 0.38/1.63 ms | 4.7 s | 95.76/−2847 |
Notes. All CNN models were trained from scratch with the same PO4AO parameters (see Table 1) and VLT 0-mag simulation environment (see Sect. 6.1 and Table 1). The Strehl and reward were calculated from the last 1000 steps of the experiment. The inference time was also calculated for VLT and ELT-scale systems, while the training time after each episode was only calculated for the VLT-scale system due to computational limitations. The corresponding integrator performance (dominated by the fitting and temporal error) for the “VLT” simulation was 93.59/−10085 (Strehl/Reward).
Current usage metrics show cumulative count of Article Views (full-text article views including HTML views, PDF and ePub downloads, according to the available data) and Abstracts Views on Vision4Press platform.
Data correspond to usage on the plateform after 2015. The current usage metrics is available 48-96 hours after online publication and is updated daily on week days.
Initial download of the metrics may take a while.