Further improvements and experiments

There are lots of directions and things that could be tried:

• More input and network engineering: the cube is a complicated thing, so simple feed-forward NNs may not be the best model. Probably, the network could greatly benefit from convolutions.
• Oscillations and instability during training might be a sign of a common RL issue with inter-step correlations. The usual approach is the target network, when we use the old version of the network to get bootstrapped values.
• The priority replay buffer might help the training speed.
• My experiments show that the samples' weighting (inversely proportional to the scramble depth) helps to get a better policy that knows how to solve slightly scrambled cubes, but might slow down the learning of deeper states. Probably, this weighting could be made adaptive to make it less aggressive in later training stages.
• Entropy loss could be added to the training to regularize our policy.
...