4. Januar 2023
The power of AI in routing and travel time prediction.
A teaser to the results of a novel machine learning approach for routing and travel time predicting. Swarm Logistics uses a proprietary machine learning service to predict travel times in routing, especially for planning. It deviates after only 4 month training on average just 2,72% from leading routing providers, who are utilizing live traffic.
Stuttgart, January 2023
Looking for pilot partner to field test the solutions.
#routing #swarmlogistics #dispatching #machinelearning #AI #artificialintelligence #tourplanning #fleetcontrol #fleetoptimization #fleetorchestration
We from Swarm Logistics have developed a new routing solution as underlying technology for our „Auto-Dispatcher“ in our Fleet Control solution and tested it over the last months. This is a short teaser and summary to the first results.
We conducted a series of tests to compare the performance of our routing service in comparison to the leading routing providers in the industry for cars, since there are much more car routing solutions available and to initially validate our new approach.
Our AI-based routing service is designed to provide commercial fleets with optimized routes and realistic Estimated Times of Arrival (ETA). One of the hardest tasks in routing is predicting the travel time correctly, because it is dependent on many external effects. We accepted the challenge.
One of the key features of our service is that it does not require live traffic data to generate routes. Instead of using live traffic data our approach utilizes machine learning techniques to predict traffic patterns and road conditions, resulting in more accurate and efficient routes. While live traffic data is important for very short routes and valuable near-term cases (minutes to approx. 1h), if used for planning in advance for some specific time in the future, from a few hours to weeks, it is not sufficient, since live traffic data is just a snapshot of a situation for a specific moment and can only give valuable information short term. The tests were done according to it, 9 leading routing providers were utilizing live traffic data with some prediction and we just predictions by our proprietary machine learning model.
In the first month we already beat most of other routing solutions in precision of expected travel duration, so here in the teaser just the results are displayed in comparison to “the” leading routing provider, who has millions of devices collecting data over decades in contrary to our 4 month of data collection.
One of the innovations of our routing service is that it takes a more generalist approach, without being trained on specific geolocations. This allows it to be more adaptable and require less data from specific origins, extrapolate and thus increase the results on a global scale much faster with less data.
The tests started in Oct. 22 with our first model and the fourth iteration in Jan .23. Over just 4 iterations and month it showed significant improvements over the initial model, with a deviation of between 3-12% in the start and decreased to between 0-7%, with only large outliers such as major traffic jams with accidents not being predicted accurately, leading to a deviation with a total average mean of 2,72%.
Especially the long hauls results are impressive. The deviations are approx. 0,2-5% with an average mean of 1,65%. To emphasize it, we are talking about e.g. 4min on 485min total travel time on 790km of distance.
While the deviation with short routes (5-20km) are approx. 4,5% average mean, which is originating in the good live traffic effects to near the planning time point test cases, to emphasize we are talking about just 1 min deviations on 25min travel time.
Overall, our AI-based routing service is a powerful tool for improving the efficiency and reliability of commercial fleet operations especially for planning.
In conclusion, our AI-based routing service is a significant advancement in the transportation industry, providing commercial fleets with highly accurate and efficient routes especially for routes at some point in the future. Its generalist approach and ability to adapt to different geolocations make it a versatile solution for a wide range of fleet needs. The impressive test results speak to the potential of this technology to revolutionize the way commercial fleets operate.
We are now starting the same tests for truck routing using the leading truck routing provider as benchmark, again the initial model and results look very promising, stay tuned for more!
Furthermore, we are looking for pilot partners for cars and trucks to field test our solutions.
Contact:
Damir Dulovic, CEO
Swarm Logistics GmbH
Koenigstr. 22, 4th Floor Marquardt Building
70173 Stuttgart, Germany
E-mail: info@swarmlogistics.net
Summary:
Bringing the average mean deviation in comparison to the leading routing provider down from over 4 model iterations.
ML V.1 – 7,25%
ML V.2 – 5,83%
ML V.3 – 4,04%
ML V.4 – 2,72%
All start and end geolocations are starting in the inner-city center, if Stuttgart, then the Swarm HQ.
All tests have been conduced on multiple times in the day, different days, in peak traffic times, non-peak traffic time, before and after peak traffic time overreaching them.
Case Descriptions:
Case1:
Stuttgart – Berlin (approx. 630km)
Test on Multi road – mainly on Autobahn and crossing multiple Autobahn, where the model data originates and overreaches to with the crossing smaller Metropolitan regions.
Case2:
Stuttgart – Stockach (approx. 160km)
Test on one Autobahn in the region containing geolocations that the model was trained on.
Case3:
Stuttgart-Hamburg (approx. 650km)
Test on a very long haul with one Autobahn crossing geolocations, where the model didn’t get trained on, crossing multiple larger metropolitan regions.
Case 4:
München-Hamburg (approx. 790km)
Test on very long hauls crossing multiple States on Autobahn with major part of driving outside a metropolitan region.
Case 5:
München-Frankfurt (approx. 400km)
Test on connecting 2 Metropolitan regions crossing multiple autobahn and states driving mainly through a metropolitan region.
Case 6:
Stuttgart-Cannstatt (approx. 6km)
Test on inner city traffic and comparison how good is the prediction for traffic estimations
Case 7:
Stuttgart-Mötzingen (approx. 50km)
Test on crossing multiple road types incl. their crossing Autobahn, Bundesstrasse, City Centers, Landstraßen, large and small city in a rural area etc.
Case 8:
Stuttgart -Esslingen (approx. 15km)
Test on 2 cities in one metropolitan area