giamman
commented
3 years ago Hi GIles, indeed if you give to the optimization routine the possibility to choose a different resolution for different parts of the detector layers, for sure it will converge to an inferior resolution near the edges than close to the center. But the gain in cost might be illusory, as the production of such detection layers would be more cumbersome and likely more difficult to industrialize. Please correct me if I am missing something.
Cheers, Andrea
Andrea Giammanco http://cp3.irmp.ucl.ac.be/Members/giammanco Phone: +32 10 473221 (UCLouvain), +32 493 581662 (mobile)
From: GilesStrong @.*** Sent: 26 May 2021 10:46 To: GilesStrong/mode_muon_tomography Cc: Subscribed Subject: [GilesStrong/mode_muon_tomography] Mechanism to optimise unused detector elements (#33)
Problem
Currently the optimisation can only adjust detector elements if a muon passes through them (and three other detector layers), however elements on the edges rarely have muons pass through them and through the three other layers, since they are liable to exit the volume. Since these elements are rarely used they should really have their resolution and efficiency decreased to save money, however the optimisation currently has no way to do this. Additionally, when the edge detector-elements are used, their resolution gets increased, even though they are rarely used because the optimisation acts solely on the current batch of muons and has no concept of usage probability.
Possible solutions (please add your own)
- Continually track the usage of each detector element, and after a warm-up period begin to decrease the resolution and efficiency of all elements according to their usage probability at each gradient-based optimisation step.
- ❌ Fix the cost of the entire detector, and uniformly rescale every element after each gradient-based optimisation step.
- Would negate the need for the cost term in the loss and mean that every update would just increase the resolution and efficiency by as much as possible whenever elements were used.
- Treat the problem as a reinforcement learning task in which actions vary the res and eff of every detector element at each step with the aim of maximising the reward of high precision and low cost.
- Twin-delayed deep deterministic policy gradient (TD3https://arxiv.org/abs/1802.09477) and soft actor-critic (SAChttps://arxiv.org/abs/1801.01290) both work for continuous actions and can use off-policy data (i.e. data gathered from previous detector states) to optimise the updates.
- When building the detector, consider only the relative values of the prescribed res and eff, rather than their absolute values.
- ❌ Remove the outer edge of the detector layers, i.e. make them 2 units less wide and long than the passive volume.
- Probably too coarse and should really be learnt by optimisation if required
— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHubhttps://github.com/GilesStrong/mode_muon_tomography/issues/33, or unsubscribehttps://github.com/notifications/unsubscribe-auth/ABFT2LQSZ3I5SEBIXHGFWGDTPSYPJANCNFSM45RP4YDA.
GilesStrong
Problem
Currently the optimisation can only adjust detector elements if a muon passes through them (and three other detector layers), however elements on the edges rarely have muons pass through them and through the three other layers, since they are liable to exit the volume. Since these elements are rarely used they should really have their resolution and efficiency decreased to save money, however the optimisation currently has no way to do this. Additionally, when the edge detector-elements are used, their resolution gets increased, even though they are rarely used because the optimisation acts solely on the current batch of muons and has no concept of usage probability.
Possible solutions (please add your own)