Repository navigation
UnboundLocalError: local variable 'simulationAction' referenced before assignment #30
Description
In this file: MCTS.py, there is such code snippet:
if Q + U > maxQU:
maxQU = Q + U
simulationAction = action
simulationEdge = edge
lg.logger_mcts.info('action with highest Q + U...%d', simulationAction)
There will be error when "if" not hold:
UnboundLocalError: local variable 'simulationAction' referenced before assignment
I encountered such error, was it a problem in game.py, or in this MCTS.py?
Activity
archcra commentedon Dec 17, 2018
I encountered this error only when set MCTS_SIMS = 1000. Does that mean, when in such case, I should change maxQU's initial value more smaller, such as -999999 ( one more 9)?
LeeTaeHoon97 commentedon Apr 10, 2022
hi archcra, have you checked your vh_loss or ph_loss?
in my case,
when i face this error, my vh_loss and ph_loss values show NaN,
so, I changed Nan values to 0
i wish you solved this problem
archcra commentedon Apr 10, 2022
tian-jie commentedon Apr 7, 2024
Can you explain more, I don't find vh_loss or ph_loss in code. I am also facing this issue now.
I tried to comment out if clause, but now it doesn't predict any more, it shows:
MCTS perceived value for O: nan
NN perceived value for O: nan
LeeTaeHoon97 commentedon Apr 7, 2024
Here is the training content I wrote down as I don't remember the details of a project I attempted a long time ago. Please understand that this translation is done using GPT.
Case1 - When initially training, errors occurred during the training process. Unsure where the problem originated, I printed out the cost values and observed that the pH values diverged in the negative direction, eventually resulting in NaN values. It appeared to be a gradient exploding problem. While I knew that conventional CNNs typically use techniques like dropout to address this, I was unsure how to handle it in a self-learning scenario. Through searching online, I found that inserting code to replace NaN values with 0 could sometimes resolve the issue. Subsequently, the code proceeded without errors. However, the loss function appeared to fluctuate erratically in the output (loss could decrease to negative values, but NaN values caused it to be converted to 0, resulting in erratic behavior). Further investigation revealed that the problem occurred during the backward pass. It turned out that the issue stemmed from a custom loss function implemented in Keras. Switching to a PyTorch loss function resolved the problem, and training commenced smoothly.
I hope it helped you.