Plateau-ing training loss and LOW train_accuracy/test_accuracy

Question

With regards to the thread topic, could anyone help to advise what is wrong with the loss_function() computation logic (especially policy_output_discrete and