DeepTraffic Leaderboard

DeepTraffic 1.0 Leaderboard

Purnawirman (74.48 mph)
Comment: "I used a single hidden layer. The improvement seems to come from taking the data as a single snapshot (time window as 0). Spent some time on hyperparameter tuning, especially L2 regularization. Submitted the model several times, because the test scores have a big variance."
Michael Gump (74.04 mph)
Comment: "I mainly played around with the L1 and L2 decay and that helped a lot. It was interesting how often the model would get stuck in suboptimal strategies and thought raising epsilon_min helped too."
Jeffrey Hu (73.59 mph)
Comment: "I preprocessed to reduce the size of the input by taking the min of consecutive cells and fed that into a three layer fully connected network. Then I tried to make gamma as large as possible while playing with the other parameters to get the network to converge."