I'm trying to solve the following problem:
- A delivery man has N packages to deliver, already sorted.
- He starts the route at
start_time, he works until
- I have historical data of his delivery schedules.
- There's no geographic information between each drop point.
How can I predict the delivery times for each package?
I've tried the following approaches, which don't seem to solve this appropriately:
- Machine learning: perfect for classifying the input into classes, which doesn't apply for this scenario. I even did the ML Coursera course in 2 weeks, just to see if any topics were suitable.
- Kalman filter: or other DSP-related algorithms focus too much on real-time, short-spanned predictions, which is not the case. No way of leveraging historical data.
- Markov chains: great for guessing the present state, not so much for a future value.
- Mathematical model: I could come up with a mathematical model (gaussian curves?) which denotes the time it takes between package delivery, depending on hour of day. Based on that, predict the delivery times.
Which topics should I focus on?
A few notes on the historical data:
- I have everything from start times, end times, and delivery times for each package. I also have all the GPS data for those routes, if that helps.
- The drop-off places should be considered random, so the delivery times look pretty random themselves.
- There's a slow-down on lunch time.