Fig. 8: Testing different optimizers.
From: Neural network approach to quasiparticle dispersions in doped antiferromagnets
a Optimization results for the t − Jz model (left) and the t − J model (right) on a 4 × 4 square lattice with t/Jz = 3, both for a single hole Nh = 1 and periodic boundaries, using stochastic gradient descent (SGD), AdaBound, Adam, Adam+Annealing and minimum-step stochastic reconfiguration (minSR), and 200 samples (1000 samples for minSR) in each variational Monte Carlo (VMC) step. All values are averages over the last 100 training steps, error bars denote the respective standard deviation. b Eigenvalues of the T-matrix (minSR algorithm80, solid lines) and of the XTX matrix (stochastic reconfiguration variant of Rende et al.81, dotted lines) before the training, for the 4 × 4t − J system with one hole and open boundaries and hidden dimensions hd = 30, 70, using 1000 samples.
