Explain the context of the issue, what is being addressed in detail.
The goal
[x] Serialize the state of the parent surface, not the one in the node itself. This is because the node represents the state AFTER the action is played, which is convenient in some aspects, and inconvenient when generating data.
[x] #66
[x] Add end turn class input and end turn should have 0,0 position
[x] Try using visits instead of exploit
[x] STATUS_CORE_ARMOR should get special treatment since it can be disabled.
[x] #72
[x] Include at least 1 guaranteed damage action in state generator (COILBLADE, WRATHSPARK, WISPSPARKS)
[x] Shuffle the dataset.
[x] Only serialize final selected nodes.
[x] Incorporate random play into state generator.
[x] Action comparison network - Output which of the two actions has higher visits in MCTS (poor performance)
Pushed back:
[ ] Fix double generation of END_TURN casts (error surpressed).
[ ] Checkpoints in notebooks.
[ ] Try using distance instead of offset vector for positions
Time tracking
Time Estimate: undefined
Time spent: 4 hours 25 minutes
Explain the context of the issue, what is being addressed in detail.
The goal
Pushed back:
Time tracking
Time Estimate:
undefined
Time spent:4 hours 25 minutes
Resources
MCTS Review of Recent Modifications and Applications https://arxiv.org/pdf/2103.04931v4.pdf#subsection.3.3 RES Blocks https://towardsdatascience.com/residual-blocks-building-blocks-of-resnet-fd90ca15d6ec