DART Overall Concept

DART will explore the applicability of a collection of data mining, machine learning and agent-
based models and algorithms to derive a data-driven trajectory prediction capability. Those
algorithms are expected to provide increased levels of accuracy while considering ATM network
effects in the prediction process, which have been rarely introduced by current state-of-the art
solutions.
The datasets used for training and testing the algorithms are a key component of the project. They
are rich high quality and extensive datasets which will be carefully crafted and curated to help
determine when the algorithms show promising results and when not. The objective is twofold:
Firstly, being able to compare different algorithms . Secondly, to identify which data sets are
relevant or irrelevant for the different learning purposes. The data included in the datasets will vary
in number and characteristics depending on the data sources selected, the granularity of the data,
the historical span, and special conditions on the ATM system when collected (i.e. severe weather, very busy sector, high demand season…). The variety of the datasets will help to discover when
each algorithm works better and which datasets are relevant. Advanced visual analytics techniques
will be used to facilitate algorithms parameters and features selection.

Once a group of learning algorithms shows promising results, they will be used for predictions in a
collaborative trajectory scenario. DART will use an agent based modeling approach for
collaborative trajectory prediction, leveraging reinforcement learning techniques to refine
predictions based on (a) potential trajectory predictions and (b) contextual information, in a
coordinated way, for groups of trajectories corresponding to agents (aircrafts).

dart-overal-concept

The output of the data-driven predictions for each trajectory in isolation will be adjusted using the
Agent-Based Modeling Module. A stream of real time trajectory information (e.g., current position,
flight plan, predictions, aircraft intent, etc.) will feed the prediction loop. The predictions for each of
these trajectories calculated with the learning algorithm(s) will feed into an Agent-Based Modeling
Module, which will allow agents to predict their trajectories based on important and recurring
patterns of situations, learning policies on how to react appropriately considering possible
interactions with others and contextual information. Doing so, we aim to allow groups of agents to
learn collaboratively on how to react appropriately on emerging situations based on their previous
experiences. This may result to new emergent phenomena to which agents should learn to react,
and so on and so forth.