Developing a simple model architecture and training it on data from #57.
Try out a few simple architectures, mostly Fully-connected DNNs, in the end a Residual Block based encoder style network was chosen due to superior performance, but significantly smaller networks perform not too far worse.
Most of the development was done in a python notebook
image of the chosen network
The goal
Describe goal of the issue, the condition for "closing".
[x] Simple model architecture using tensorflow.
[x] Define a game-state serialization format and write a generator for it in gdextension.
[x] #57
Time tracking
Time Estimate: 5 hours 0 minutes
Time spent: 7 hours 10 minutes
Developing a simple model architecture and training it on data from #57.
Try out a few simple architectures, mostly Fully-connected DNNs, in the end a Residual Block based encoder style network was chosen due to superior performance, but significantly smaller networks perform not too far worse.
Most of the development was done in a python notebook
The goal
Describe goal of the issue, the condition for "closing".
Time tracking
Time Estimate:
5 hours 0 minutesTime spent:7 hours 10 minutesResources