AI Case Study

The researchers posit that their "approach has the potential to address the limitations of the existing hand-tuned heuristics. In particular, our approach learns to optimize packet classification for a given set of rules and objective, can easily incorporate pre-engineered heuristics to leverage their domain knowledge, and does so with little human involvement.

[The] solution uses deep reinforcement learning (RL) to
build efficient decision trees. There are three characteristics
that makes RL a particularly good fit for packet classification.
First, the natural solution to build a decision tree is to start
with one node and recursively split (cut) it. Unfortunately,
this kind of approach does not have a greedy solution. When
making a decision to cut a node, we do not know whether
that decision was a good one (i.e., whether it leads to an
efficient tree) before we finish building the actual tree. RL
naturally captures this characteristic as it does not assume
that the impact of a given decision on the performance objective is known immediately. Second, unlike existing heuristics
which take actions that are only loosely related to the performance objective, the explicit goal of an RL algorithm is to
directly maximize the performance objective. Third, unlike
other RL domains such as as robotics, for our problem it is
possible to evaluate an RL model quickly (i.e., a few seconds
of CPU time). This alleviates one of the main drawbacks of
RL algorithms: the non-trivial learning time due to the need
to evaluate a large number of models to find a good solution.
By being able to evaluate each model quickly (and, as we will
see, in parallel) we significantly reduce the learning time.

To this end, we design NeuroCuts, a deep RL solution for
packet classification that learns to build efficient decision
trees."