How Can Artificial Intelligence Speed Up the Journey from Dakhla to Paris?

The roads linking Dakhla in Morocco to Paris in France are a real laboratory for modern logistics. Researchers gathered thousands of data points on traffic, weather, speed limits, and fuel costs along this route. They then tested various machine‑learning algorithms, from classic linear regressors to sophisticated neural networks. The outcome? Neural networks identified routes that cut travel time by 12 % and costs by 9 % compared to traditional methods, while keeping safety standards high. The secret lies in these models' ability to analyze many variables at once and adapt to sudden changes, like a jam or a storm. Moreover, the system respects European transport regulations, avoiding roads with weight limits or night‑time bans. In practice, a truck equipped with this “digital brain” can pick the smartest path in real time, saving time, money, and reducing emissions. This study shows that with high‑quality data and advanced tech, road logistics can become more efficient and sustainable, paving the way for a smarter transport future.

The dataset built for the Dakhla‑Paris corridor includes over 150,000 records, each containing variables such as average speed, traffic density, congestion index, weather conditions (precipitation, visibility, temperature), weight restrictions, and time‑of‑day bans. During preprocessing, researchers normalized values, handled missing data with k‑nearest neighbours imputation, and engineered features like a "predicted waiting time" by combining historical incident data and weather forecasts. Tested models comprised linear regression, Random Forest, Gradient Boosting, and deep neural networks (DNN) with three hidden layers of 128, 64, and 32 neurons, ReLU activation, and Adam optimization. The loss function combined weighted travel time, fuel cost, and safety penalties to reflect operational priorities. The DNN achieved an R² of 0.92, outperforming other algorithms (Random Forest R² = 0.86). Five‑fold cross‑validation confirmed model robustness, with an error standard deviation below 3 %. Additionally, a constraint‑satisfaction module checks real‑time compliance with EU regulations, discarding non‑conforming routes. The entire workflow ran on a cloud platform, leveraging GPUs for neural network training, cutting training time from 12 hours to 45 minutes. These results show that integrating high‑quality data, targeted feature engineering, and deep‑learning models can produce significant route optimizations, yielding measurable economic and environmental benefits.