IEEE Computational Intelligence Magazine - November 2019 - 24

network hardware and acts as a centralized plane
in which an AI agent can interact with the whole
network to generate an optimal strategy. However, while the advantages of centralized optimization are clear, the overhead of closed-loop
control implemented through a centralized AI is
high. This overhead includes not only the communication overhead for receiving and sending a
large amount of data but also the computational
overhead on the AI agent side for training and
execution. In the centralized paradigm, all routers need to be
programmed to build a single flow-forwarding path. In addition, every time the network status changes, the controller
needs to recompute the forwarding logic. For large-scale networks with ultrahigh dynamics (on the millisecond scale),
this excessive communication pressure and high computational burden are unacceptable.
As discussed above, the centralized and distributed paradigms, as the two ends of a spectrum, are both imperfect and
have corresponding advantages and disadvantages. A distributed
architecture carries the risk of non-convergence of the learning
process, but it offers faster forwarding and processing speeds for
each packet. In contrast, completely centralized learning is
advantageous for global optimization but may incur excessive
overheads in terms of communication and computation.
Therefore, from our perspective, the centralized and distributed
approaches should be treated as complementary rather than
mutually exclusive. In this paper, as shown in Fig. 2, we propose a hybrid AI-driven control architecture that combines a
"network mind" (centralized intelligence) with "AI routers"
(distributed intelligence) to support different network services.
Before we detail the operations in our AI-based routing
paradigm, let us start by reviewing the current state of development of routing protocols. Early on, the IP protocol won
the battle between connectionless and connection-oriented
routing and between source routing and distributed routing.
As shown in Fig. 3, in the IP protocol, each router establishes a routing table based on its local information and communications. This routing table contains the next-hop node
and a cost metric for each destination. Based on this hop-byhop forwarding paradigm, a data packet needs to carry only its
destination address in its header, which is beneficial for network scalability and robustness. However, because of the connectionless and distributed characteristics of the IP protocol,
traditional IP routing provides poor support for traffic engineering and QoS guarantees. To support high-QoS (highbandwidth, delay-sensitive) services, connection-oriented and
source routing mechanisms have begun to receive attention
once again. For example, as shown in Fig. 4, Multiprotocol
Label Switching (MPLS) uses connection-oriented label
switching and explicit paths (source routing) to establish temporary network tunnels between senders and receivers. This
predetermined temporary tunnel routing strategy provides an
easier and more efficient QoS-guarantee mechanism for service providers. However, full-mesh network tunneling is

In an inherently distributed system, state
synchronization among all routers is extremely
difficult, especially with increasing network size,
speed, and load. With the development of SDN
technology, centralized AI-driven routing strategies
have received considerable attention.
construct a closed-loop network control paradigm. As illustrated in Fig. 1, our paradigm consists of three layers, called
the forwarding plane, the awareness plane, and the intelligent
control plane.
The forwarding plane is responsible for forwarding data packets from one interface to another in distributed network equipment. Its operation logic relies completely on the forwarding
table and configuration instructions issued by the control plane.
The purpose of the awareness plane is to monitor the network status and upload the results to the control plane. Network monitoring and awareness are prerequisites for ML-based
control and optimization. Therefore, we abstract this new layer
called the awareness plane for the collection and processing
of monitoring data (for tasks such as network device monitoring and network traffic identification) to provide network status information.
The intelligent control plane is responsible for feeding
control decisions to the forwarding plane. The AI&ML-based
algorithms are deployed in this plane to transform the current
and historical operation data into control policies.
These three abstract planes together constitute a closed-loop
framework for AI&ML deployment in networking. In analogy
to the human learning process, the forwarding plane acts as the
"subject of action," the awareness plane acts as the "subject of
observation", and the intelligent control plane acts as the "subject of learning/judgment." Based on these three planes for
closed-loop control, an AI&ML agent can continuously learn
and optimize network control and management strategies by
interacting with the underlying network.
3.2. Centralized vs. Distributed

As described above, a three-tier logical architecture is proposed.
However, when this abstract logical concept is deployed in a
real-world network, the placement of the intelligent control
plane (centralized or distributed) is critical to the efficient operation of AI-driven networking.
How far away the control plane can be located from the
data plane has long been a controversial topic. In traditional
distributed networking equipment, the control plane and the
forwarding plane are closely coupled. Each node has only a
partial view of, and partial control over, the complete network.
When AI&ML-based algorithms are applied in such a network,
the learning process will suffer from severe non-convergence,
particularly when a global optimum is sought. In contrast, in an
SDN architecture, the control plane is decoupled from the

IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE | NOVEMBER 2019

IEEE Computational Intelligence Magazine - November 2019

Table of Contents for the Digital Edition of IEEE Computational Intelligence Magazine - November 2019