# Project Ideas

## Contents
* [VAE analysis of a phase diagram](#vae-pd)
* [Forward prediction of the Overhauser field](#overhauser)
* [RL Control of a Quantum Cartpole](#quantum-cartpole)
* [Self-play optimization for error decoding](#self-play)
* [Quantum error correction using a value network only](value)
* [A new type of RL controller for quantum physics](#qp-rl)
* [Capsule networks for the Ising model transition](#capsule)
* [An AI for quantum games](#qttt-ai)
* [Automatic action discovery for RL agent](#autodiscovery)

---

### VAE Analysis of a phase diagram
* Choose two or multiple models with varying numbers of Hamiltonian parameters ($N$) and unique phases ($M$).
* Numerically obtain groundstates sampled from the phase diagram (exact diagonalization, matrix product states)
* Use a variational autoencoder (VAE) to compress the samples down to a latent space with a dimension $D$
* Hypothesis: $D \sim M$ and *not* $D \sim N$.

### Forward prediction of the Overhauser field
In spin-qubits, the nuclear spins in the environment cause a background field called the 'Overhauser' field. If this field can be estimated/predicted, it can be _used_ to implement universal control on multiple qubits.
Estimating the field can be done using Bayesian estimation, which can also be made adaptive. Let's investigate this, and *use* it in actual experiments.
#### Refs
* https://arxiv.org/abs/1102.3700
* Henrik Bluhm paper with Yacobi

### Reinforcement learned control of a quantum cartpole
The classical cartpole is the standard benchmark for reinforcement learning (https://gym.openai.com/envs/CartPole-v1/), and this project will investigate a quantum version of it. A previous result using DQNs is available, but with some unclear choices (https://arxiv.org/abs/1910.09200).
This project could well be combined with "A new type of RL control".

### Self-play optimization of quantum error correction
TODO

### Quantum error correction using a value network only
Instead of using a neural network that represents a Q-function, we will only use one that represents the value-function. This should be much faster. Together with Mats Granath at Chalmers (Sweden).

### A new type of RL control for quantum
Rough idea: Optimize a _sequence_ of gate voltages that produces the 'best' quantum point contact or FQHE plateau. Instead of suggesting the voltages directly, the RL agent must choose between actions of type '+dV', '-dV' and 'done'.

### Capsule network for phase transitions
Rough idea: Investigate the use of capsule networks for phase classification. Based purely on the idea that capsnets: "... can identify an object independent of its pose, rather than having to learn to recognize the object while including its spatial relationships as part of the object" (https://en.wikipedia.org/wiki/Capsule_neural_network)

### AI for Quantum TiqTaqToe
TODO (also, more of a computer science project)