Added rationality penalties.

KEY FEATURES

• Added a differentiable rationality penalty so that trips are sequential.
• Slightly better performance

Almost ready for production..

Key Features

• Using Duration instead of TripEnd - much more sensible values.
• Feeding back softmaxed trip purpose values rather than argmaxed, for differentiability.
• Losses are matrix-based - calculated once a full travel week has been populated.

TravNet v0.0.1

KEY FEATURES

• Single layer, 9 unit RNN.
• 1-day ahead and 7-day ahead recursive forecasts for travel diary
• Penalty for predicting $TripEnd_t < TripStart_t$
• Partially decreasing loss