Commit 5522472e authored by Vivien Kraus's avatar Vivien Kraus
Browse files

Reduce the learning rate

I often get times where all the predictions are set to 0.  This is
because in the beginning where there is a lot of exploration, the
regression is simply based on the bid and whether we are the taker.
So, most parts of the network are "discarded", which is to say the
weights optimize so that the hidden neuron is never activated.
However, at the end of the exploration, we need to exploit these
neurons.
parent a70be33f
......@@ -62,7 +62,7 @@ perceptron_construct_static_default (TarotPerceptron * perceptron)
};
size_t n_hidden_layers = sizeof (hidden_sizes) / sizeof (hidden_sizes[0]);
size_t n_weights = sizeof (weights) / sizeof (weights[0]);
perceptron_construct (perceptron, n_hidden_layers, hidden_sizes, 1e-6);
perceptron_construct (perceptron, n_hidden_layers, hidden_sizes, 1e-8);
perceptron_load (perceptron, 0, n_weights, weights);
}
......
......@@ -75,7 +75,7 @@ parse_options (int argc, char *argv[])
uint8_t *seed;
char *end;
exploration = 0.05;
learning_rate = 1e-7;
learning_rate = 1e-8;
yarrow256_init (&generator, 0, NULL);
yarrow256_seed (&generator, strlen (default_seed), (void *) default_seed);
while (1)
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment