Fig. 4
From: Multiparameter optimisation of a magneto-optical trap using deep learning
Cost landscapes and convergence predicted cost landscape cross sections generated by the model after exploration of the parameter space. Each 1-dimensional slice (b) is generated by varying each parameter independently over the available range while keeping other parameter values constant at their best known value shown in a. The red and orange curves represent arbitrarily chosen points that demonstrate landscapes for intermediate values and boundary limited values respectively. The blue curves represent every other parameter not highlighted in the experimental run. c The convergence of the model is attained by observing the measured and predicted costs as shown in the top plot. The red shaded area corresponds to experimental noise. The middle plot shows the scaled difference between these two measurements and the associated moving average as the SANN explores different regions of the cost landscape. The lower plot shows the distance of a given parameter set from the best observed parameter set
