I'm quite new to TF. I have written a Deep Q-Learning CNN to control a simple driving simulator. I've managed to plot the weights, biases and outputs of my fully connected layers on TensorBoard however, I'm not sure what I'm looking for. I've watched a youtube video that mentions a normally distributed set of weights is good. I was wondering if anyone has any other debugging tips. Also, I'm using Relu for my activation functions but many of my 'rewards' and hence my 'qualities' output by the CNN are negative should I be using something like a Sigmoid instead? I'll grab some screenshots of my TensorBoard and post them too.
Asked
Active
Viewed 287 times
2
Maxim
- 52,561
- 27
- 155
- 209
Joshua Patterson
- 439
- 5
- 12
-
UPDATE: I've changed the activation functions to Sigmoid as (from ReLu). I'll post the Histograms when the have some time to settle down. I also found some links [here](https://www.tensorflow.org/tutorials/mnist/tf/) and [here](https://wookayin.github.io/TensorflowKR-2016-talk-debugging/#16) that I'm reading now. I'll read and comment again later. – Joshua Patterson Feb 03 '17 at 03:39