Neural net to approximate the simulator
Inputs either:
recurrent & 72 current numbers
72 numbers & first difference & second difference

UCB rollout policy takes random actions. It would be cool to use the sort-of-bad-but-better-than-random neural network controller as the rollout controller. At very least, anything better than random.

Since switching to the yaml configuration model, using the same model at the same checkpoint for evaluation and control became much harder. I don't actually know a great way to do this, but let's figure it out.

TreeWorker pooling was worthwhile when each game had to be on its own classloader. Now that this is no longer the case with GameUnified. The complexity added by using pooling is likely not worth it.

The links are flying everywhere!

Panel which monitors different threads in the tree UI doesn't work with the new yaml system. Fix at some point.

We might want to run a bunch of games as part of the rollout evaluation. This can be done in a hacky way inside the sampler itself, but I'm not fully satisfied.

It would be nice to be able to propagate values back along the tree in multiple ways.
Candidates:
Best overall
Average (current strategy)
Best of three with rollouts or value function evaluations

Spent a long time tracking down inconsistencies in controller results only to find that the state normalization method was hardcoded differently now than it had been a month ago. Make this part of the configuration to keep it more obvious.

Make it so GameLoaders can be saved and loaded to/from file.

With the UCB sampler, workers tend to deadlock when there are a lot of workers. I think that some sections of the tree might be being locked from other workers too much. When one worker is expanding from a node, it doesn't allow other works to also try another potential action from that node. This should be changed.

Lots of logging spams up the console. Also on the CI server, there are graphics errors that somehow don't cause the build to fail, but still seem like bad things.

Broke sometime after adding the built-in log console.

For most predictive controllers, the state may be set by setting the 72 visible state variables (approximate) or by restoring the serialized game (exact). Make this selectable and obvious.