Given that
machine learning essentially deals with developing algorithms and systems
which “learn” from data, it is not surprising that machine learning has a great
deal of applications in the development of artificial intelligence. Being a
science fiction fan, it is quite possibly a bridge to autonomous, bipedal,
humanoid, robots which hopefully won’t end up like terminator.
Here is a
lecture from UC Berkeley from the electrical and computer engineering
department from UC Berkeley.
At the beginning of
the lecture, he starts off by explaining an experiment in which he attempts to
train a remote control model (not the real big kind) helicopter to perform some
rather intricate maneuvers which can normally only be done by expertly skilled
model helicopter pilot.
The data collected, as I imagine being the training set for
this scenario is the position, velocity, orientation and the angular rate of
the helicopter. They also collected data on the controller inputs. He shows
data on the trajectories of various iterations of the same helicopter air show
in order to gather data from that in order to apply the machine learning, also.
Interestingly, he uses the Needleman Wunsch sequencing alignment algorithm
across the iterations of the human trial trajectories in order to get a
learning trajectory I imagine he will be feeding to the machine learning
algorithm (I discussed Needleman Wunsch in one of my previous post, but it had
to deal with bioinformatics, and not AI).
He mentions that the primary purpose of machine learning in
this context is to train a computer to do the things which is generally
associated with the same learning a human pilot would undergo, (I.e., a human
pilot spends many hours training and doing the same maneuvers until it becomes
“muscle memory” and he may know exactly what some action on the controls will
do before he does them.)
Around 27:46 in the video, he begins
lecturing on how machine learning can be used to train robots to perform
surgery. The surgery however is not terribly crazy like a brain transplant. It’s
just a basic surgical know tie. At the end of the video, he goes into
demonstrating how he applied machine learning to teach a quad pedal robot how
to walk across a rough surface. Just check out the video.
The lecturer in the video is
Pieter Abbeel, Department of Electrical Engineering and Computer Sciences, UC Berkeley
Pieter Abbeel, Department of Electrical Engineering and Computer Sciences, UC Berkeley
This is a pretty interesting application of machine learning. Berkeley seems to be on the cutting edge of interesting applications to machine learning, as a couple years ago they used machine learning to make the best AI StarCraft player. I think both of these could have pretty immense impact on the future, because both of the things that they are teaching computers to do seem simple but are actually extremely complex. What StarCraft represents is the reaction to situations, something that would be extremely useful in a use case such as he discusses with the UAV. Responding quickly to an immense number of inputs is very important when it comes to autonomous AIs and something that they seem to be trying to tackle.
ReplyDeleteA major difference in these two studies is that, as he discusses around 16:45, the act of successfully completing these maneuvers is actually fairly repeatable. With the StarCraft AI however, the opponents actions are extremely unpredictable, meaning that the AI must respond to opponent actions that it cannot predict at all.