- established two phase transitions in the control problem believed to be generic: from an overconstrained phase to a glassy phase at $T_\ast$, and from the glassy phase to an underconstrained phase at $T_{min}$. Both have simple interpretation on Bloch sphere. - fixed a tentative look for the plots of the 2LS part of paper - in the appendix go: scaling for even and odd effects, dynamical phase transition in ramp time, entropy, comparison to LZ (no CD driving), continuous protocols, inst fidelity, etc.

- What order parameter to compute exactly (fluctuations in h_i^t) - Kink and phase transition - Entropy and fidelity histograms would be helpful

- Analyzing data for SGD (should have everything we need). Plots on dropbox - Recording available data in excel-like sheet on dropbox (using Numbers)

- Running jobs for dt=0.01 and 0.005 for bang-bang8 and continuous protocols. - Running jobs with fixing number of time steps for varying Tmax.

- Conda install not working, need to clone repo and use python setup.py install.

- Discussion with Marin. - Producing plots/ running jobs on the scc

- Discussion with Marin. - Running continuous protocol and checking vs. bang-bang protocols

- Compiled stochastic gradient ascent data - Fidelity vs time plots - Fluctuations in fidelity - Computation time (time needed to converge) vs evolution time - Edward-Anderson order parameter - vi-SNE plots - Finished reviewing/commenting draft

- First running zero-temperature with various evolution times - Saving protocols and all internal info for later use (computing order parameters and different metrics, etc.) - Quenching from T=0.04 with different number of quenches - Preparing slides - Implemented gradient stochastic descent (zero-temperature annealing, stops automatically when local minima is reached) - Made the code easy to use from the command line and added multiple utility functions to code - (Alex) Reviewing draft - (Alex) Running jobs on local cluster (!)

1. The problem has a phase transition to a glass in time at T': think of protocol as a spin configuration (Ising --> Potts for bang --> cont). Check this using the annealer at zero temperature by mapping out the Edwards-Anderson order parameter.
   

2. Dries: map out the fidelity DOS, and define a temperature (this is the RL and annealing temp!). This temperature can be negative (deep consequences for the learning problem).
   

3. Toli: keep the norm of the total magnetic field fixed (--> rescale coeff of S^z by \sqrt{1+h^2(t)}). Expecting the continuous protocol to give better fidelities than the bang-bang one.
   

4. Pankaj: optimisation problem is a glass -- if the reward is given at the end of the protocol, the effective model should be a random-energy model; if rewards are given at each step, the `energies' (i.e. fidelities) are correlated [no random-energy model] but it's still a glass.
   

5. Pankaj: should use the fidelity DOS, an estimate of which is updated online, to choose the annealer/RL agent's exploration temperature.
   

6. Toli: as a many-body example take the spin model and ramp thru a critical point and without a critical point. Map out best fidelity as a function of L and T. Compare annealer, RL agent and counter-diabatic (?) protocols.

1. wrote automated plot functions
2. created plots and Bloch sphere movies for bang-bang and continuous protocols

1. Simulated annealing can now be constrained to follow the same set of actions as the reinforcement learning method
2. This is an option that can be changed, so SA also is unconstrained. 3. Added a note on simulated annealing on the dropbox

1. tried out different exploration-exploitation schemes
2. reducing Sent and Energy doesn't work (used only 20 t-steps for L=4 with symmetries)
3. wrote a data evaluation code to read in the data files and updated existing code

1. plots fidelity, energy difference (above inst GS), energy fluctuations (in inst H(t)), entanglement entropy (for L>1), and diagonal entropy in basis of target state
2. producing simulation for vector on Bloch sphere (requires "sudo port install ffmpeg")

1. alternating episodes following the softmax Q and best encountered policies improves policy fast in the beginning but still has problems after replays are turned off

1. It appears the 2LS problem might just be convex (just do stochastic gradient descent, no need for annealing)
2. New optional parameter: Number of times the fidelity is evaluated. This fixes the computational complexity so one can compare different methods.
3. prepared RL code for cythonisation.

1. improved GD in h, and traces update
2. introduced search_sorted in function find_tile to improve speed [binary search ~log(N), argmin -- ~N].

1. cleared up the code in RL_2LS from junk
2. introduced softmax exploration 3. restricting actions depending on state; eliminated 'wandering off the grid'

1. params are: h_i=-1=-h_f, h(t)\in[-2,2]
2. introduced different branches for annealing and RL
3. simulated annealing matches RL

1. ramp velocity programmed in a tabular manner
2. exploration slowed down due to increase of RL state space