Things I am actively working on are kept here- subject to frequent change
Projects which do not yet have results have no page link, naturally
My largest and most voluminous project is the investigation of a machine learning algorithm I developed in graduate school. It is designed to be able to learn dynamic sequential solutions and responds to a variety of novel training methods, making it very flexible. Currently, it can learn to run mazes, solve the Tower of Hanoi, Navigate turn-only grids, addition, and a number of other problem-solving tasks. Currently examining stochastic predictive properties as well.
Inspired by SpaceX's attempts to soft land a booster, I've been tempted into playing with rockets. I want to launch one, then use aerodynamically guided ballistic descent to guide it to a destination, using inertial AHRS, GPS, and possibly a radio beacon. Once at the destination, I want it to use nose-mounted motors to slow its descent to the survivability region. Right now, I'm testing instrumentation- I verified in a non-flight test that an infrared sensor can detect ignition and cutoff. The next hold-down tests will be in-flight ignition of braking motors and then use of the ejection charge to cause the initial roll to descent profile. We'll see how it goes (5-1 against survival).
It has only been recently that I've been exposed to deep learning and the methods associated with it, but I'm already entranced by the possibilities they offer. Currently, I'm exploring state space reduction with autoencoders, reinforcement strategies for multilayer RBM networks, and using associative mapping for applying computer vision to local processing tasks (such as obstacle avoidance).
|
It occurred to me that it would be possible to use modified implementations of Dijkstra's algorithm to find optimal paths through graphs whose edge 'distances' were any metrics,and thus could be used to find, say, a greatest-likelihood sequence of actions to achieve a goal if the transition probabilities for the possible state-action pairs were known. Since other metrics can be applied as well, this method can be used for path planning similar to STRIPS, but with a probabilistic component, which allows for both learning of a world model via experimentation (populating an SxSxA matrix) and construction of a decision tree analog without the existence of a hierarchical model. So far, I've used this method to teach navigation on a grid, and to implement the traditional STRIPS toy problem in simulation (in room-open door- get soda- return). The agent has performed successfully in both contexts.
I have been interested in fuel cells, particularly those using hydrogen as fuel, and the production of the hydrogen for those cells. In particular, I think it's morbidly ironic that the bulk majority of industrial hydrogen comes from fossil fuel processing, so I've been experimenting with generation methods. I'm most intrigued by the Lye-Aluminum-Water method, because most references simply treat it as a science-project grade technique, due to the hazards of using Lye. I'm working on a small reactor (pictured above) using active temperature regulation to stabilize operation of the reactor.
|
