对应版本2.9
优化以下错误提示：UserWarning: Creating a tensor from a list of numpy.ndarrays is extremely slow. Please consider converting the list to a single numpy.ndarray with numpy.array() before converting to a tensor.
Change Critic NN as Multi-NN
wrong remain Time Fix
wrong remain Time Fix, what a stupid mistake...
and fix doubled WANDB writer
Deeper TargetNN
deeper target NN and will get target state while receive hidden layer's output.
Change Middle input
let every thing expect raycast input to target network.
Change Activation function to Tanh
Change Activation function to Tanh, and it's works a little bit better than before.
save training dataset by it target type.
while training NN use single target training set to backward NN.
this improve at least 20 times faster than last update!
while game over add remaintime/15 to every step's rewards. to improve this round's training weight.
fix get target from states still using onehot decoder bug.
add discrete and continuous action in same NN model.
model save and load.
reward is increasing, converge was observed.
this two models are seems good:
Aimbot_9331_1667423213_hybrid_train2
Aimbot_9331_1667389873_hybrid
add GAIL GAILMem GAILConfig Class.
add HumanAction record to save expert data.
add tackState future for stack multiple states to let agent knows what happened before.
Unity:
No more detect Closest enemy info. Add different density sensor let agent get more state information on the center of view.
Adjust Start Scene UI manager. Add in game visible rayCast & information that rayCast detect.
Python:
Start use mypy black and flake8 to format Python.
