Closed goodnamegood closed 6 years ago
For the games we tested on, elu
got slightly higher scores than relu
. But universe-starter-agent is intended as a starting point, a framework for experimentation, so you should experiment with that and other modifications to see what gets the best results.
As titled. I think this line should be x = tf.nn.relu(conv2d(x, 32, "l{}".format(i + 1), [3, 3], [2, 2])) instead of x = tf.nn.elu(conv2d(x, 32, "l{}".format(i + 1), [3, 3], [2, 2]))