The field of AI has expanded immensely, evolving from a highly technical discipline into common
parlance. It's crucial to clarify what Artificial Intelligence means and how it achieves its goals.
To maintain intuition, let's consider the illustration in the left figure. In essence, AI's purpose
is to enable a computer to "understand" a picture, similarly to how a human would.
With our goals established, we can now explore how this "understanding" is achieved. At this stage,
we remain at an intuitive level without delving into technical considerations. For a computer to
"understand" a picture, the path taken by AI is multi-stage, starting from lower-level tasks and
progressing towards higher levels. The first layer is illustrated
.
The "first" layer involves the identification of items (objects, animals, humans, etc.).
Technically,
this isn't truly the first layer, as it relies on lower-level layers. One might imagine that a dog
is identified as "an item with a somewhat glossy skin, four legs, etc." What's crucial to understand
is that the identification of items relies on patterns: a dog is an item with certain
properties, a girl is a human being with certain properties, and so on.
The same logic applies to the
:
The next layer could be the identification of movements, once again with the help of associated
patterns: the position of the dog implies that it's "about to jump," derived by specific patterns
("if the dog is seen standing and crouching, then...").
Finally, higher-level layers involve more "hidden" interpretations, such as the intents of
the
action (the dog's action is a jump, interpreted further as a jump to catch a treat) and even the
feelings (the girl's smile is interpreted as "playing with her dog and having fun in the game").
Going further, we have empathy (the girl plays and offers a treat because she loves
her
dog).
Without delving into technical details, let's introduce, in an illustrative manner, how AI works. The
framework for implementing image recognition is called neural network, and this technique has
a long history, starting in the early 1940s. Successful implementation for fields such as natural
language or image recognition became possible with a certain framework, denoted convolutional neural
networks. Convolutional neural networks can be understood intuitively, and it's crucial
to grasp them to understand the meaning of deep in the deep learning expression.
Here, we introduce, without technical detail, what convolutional neural networks are and how they work.
Consider a picture, consisting of a grid of pixels, illustrated on the left. Our goal is to extract
lowest level patterns for each pixel of the picture. In other words, we want to know if a
given pixel is part of a small straight line, a curved line, a small circle, etc. The knowledge of
these lowest level patterns will allow us to start understanding the picture by applying
rules on these lower level patterns, like identifying a hand once we have identified five fingers
and a palm.
The identification of lowest level patterns relies on a sequence of tricks that we'll only
survey. The first trick is to consider a with a cruder resolution. Each square of this cruder grid captures a
corresponding portion of the initial image, as
. Now,
without entering into the details, we assume that we have developed a process
determining if the square of the initial image contains a given pattern, and this information is
stored in the square of the cruder grid, as
.
The process illustrated on the figure is a combination of processes that we do not detail.
What's important is obtaining information on a cruder grid about a given pattern. Once done for a
pattern, we can reproduce it for any lowest level pattern, as
. So, we end up
with a set of cruder grids, each one containing information on whether a given pattern is in the
corresponding big square of the initial picture.
At this step, we are nearly done, and we can consider developing higher level patterns by
playing with the smaller grids, with rules such as: "If I have pattern A in a
square and pattern B in a neighboring square, then I identify a higher level pattern on
these sets of squares". In other words, from the lowest level pattern, we can imagine how
to design higher level patterns one step above, and from these, how to proceed again one step
above, and so on.
When delving into the realm of Artificial Intelligence (AI), a pivotal aspect is the utilization of
patterns that are intricately layered from lower to higher levels, progressively unveiling
more nuanced meanings. As illustrated in the earlier example, the lower-level patterns are somewhat
rudimentary, involving the identification of specific aspects of an object to derive its
"identity". However, the higher-level patterns are remarkably refined, enabling the modeling of
sophisticated concepts like empathy.
In the field of Artificial Intelligence, the current forefront achievement is within the domain of
deep learning. It is noteworthy that this technique, while not entirely novel, is conceptually
straightforward, especially in its initial application, the image recognition. The initial method
employed a neural network, using a strategic approach of a "multi-layered" network aptly suited for
discerning patterns like lines and curves. Without delving into technical intricacies, this neural
network was termed convolutional, harnessing the efficiency of the underlying convergence
algorithm for neural networks, referred to as backpropagation, which surprisingly demonstrated
swift performance.
The key takeaway is that a well-structured neural network (specifically, the convolutional neural
network) stimulates the machine to identify patterns, facilitated by a multi-layered
architecture dedicated to recognizing lower-level patterns. This technique gained widespread acclaim for
achieving remarkable performance, notably in games such as Chess or Go (refer to AlphaGo and Alpha zero). As
elucidated in the provided links, the depth of the computer's gameplay has influenced top-tier chess
players to adopt key concepts from this style of play.
Chess and Go share commonalities, including a combinatorial explosion of possibilities (excluding brute
force exploration of all possibilities by an algorithm) and reliance on patterns for gaining 'slight'
advantages, effective in the long term. Hence, an expert player must cultivate a profound understanding
of these patterns, coupled with the ability for short-term computations to "utilize" these patterns in
the long term without succumbing to tactical maneuvers.
Our approach to optimizing Air Operations shares notable similarities with deep learning. We recall that the inception of deep learning relied on a clever application of mathematics, particularly
the convolutional neural network, enabling the identification of numerous low-level patterns at minimal
computational cost. Similarly, our methodology harnesses a judicious use of mathematics, facilitating the
implementation of specific and cost-effective algorithms for detecting lower-level patterns in
air traffic. These lower-level patterns encompass various dimensions of aircraft maneuvers, including
geometric aspects and supplementary fields used to identify aircraft performance. Higher-level patterns,
derived from these maneuvers, encompass the Instructions issued by air traffic controllers, which we
also identify. Moving to a more advanced level, we delve into the strategies employed by
controllers, along with associated Workload. Progressing further, we evaluate the operational
risk, assessing how these strategies may falter and potential undesirable outcomes.
A noteworthy distinction from Deep Learning is that we do not require pre-training our system
with a training data set. As we will explore later, the rules we identify in our patterns arise from the
operational context combined with expert knowledge. Additional rules can potentially be incorporated to
enhance the existing patterns with additional features, such as descriptive parameters.
Therefore, our methodology can be tailored to any environment, aiming to enhance Air Operations by
reducing fuel consumption and minimizing operational risk.