Abstract
The goal of the project is to estimate the fleet size, creating models and optimizing parameters to get the best revenue and utilization of the vehicles on road.
The Autofleet’s platform is a tool which provides elastic supply of vehicles to any source of demand.
There were many projects approached during the project:
- Simulator optimization: Find the optimal parameters of the fleet simulator, to optimize utilization of the vehicles and completion rate of trips.
- Time series analysis.
- Predict number of trips by hour by tile (~neighborhood)
- Analyze clients reviews sentiment by different categories.
Challenges
- The most difficult challenge was the lack of data. In some cases synthetic data was created to help.
- On the NLP project, the fact that the data wasn’t labeled was a challenge, but approaching it as an unsupervised sentiment and with some libraries we achieve a decent naive result.
Achievements (according to KPIs)
- Find the amount, time, distance distributions of the trips by hour during the year.
- Determine that a tile level of prediction is not worth it to do, because most of the trips are located in the city center, and the distances are big, so the losses are greater than the profits.
- Acceptable unsupervised sentiment analysis by 7 categories per review.
Further development
- Integrate the sentiment analysis with google API to increase the accuracy. (NLP)
- Label more data for getting more metrics. (NLP)
- Define more grammatical rules. (NLP)
- For fleet optimization collect data to to good prediction.
