The site horsehour.com was registered on GoDaddy in March 2013, but went online until May 2017. To record some of my random thoughts on research and programming, this Jekyll-powered, Github Pages hosted blog was setup. I prefer taking notes with $\rm \LaTeX$, because of its excellent rendering quality of complex mathematical equations. Therefore, the posts are written in a hybrid of Markdown (for text) and $\rm \LaTeX$ (for equations) and converted to HTML using Pandoc. Sometimes, pdf2htmlEX is also used to convert PDFs to HTML documents with the original layouts of the PDF docuemnts well-preserved.

Reinforcement Learning based Voting Winner Determination Procedure

All important signals, including the current and visited states, the depth, the size of alternatives never been a winner, the number of winners, the percentage of winners and non-winners in the state, the predicted probability based on other features, are included in a new model that learned from the experience (episodes from root to leaf node). By learning such kind of model, we can get rid of the cache checking and pruning component, we choose the path according to the prediction results from this enhanced function to reduce the number of nodes to extend.

Some driving behaviors may have dramatic impacts on the fuel consumption. We are investigating the eco-driving strategy from speed profiles in GPS trajectories.

Speed Prediction from Taxi Trajectory Data

We are given a trajectory data of DiDi Express and DiDi Premier drivers within the Second Ring Road of Xi’An City. All track points are bound to physical roads with resolution about 2-4 seconds. The problem is to predict the average speed of all vehicles running on a road segment (either north- or south-bound) at specific timestamp.

A Short Review of Network Science

Network science, has its root in graph theory, is evolving to be a multidisciplinary research field. It studies the network representations of physical, biological, and man-made systems, and designs models to reproduce and predict them. One key characteristic of a complex network is its degree distribution $P_k$, i.e, the probability that a randomly selected node has $k$ links. Based on degree distribution, many complex networks in real world, including communication networks, transportation networks, Internet, social networks and biological networks are characterized by a power-law degree distribution $P_k=Ck^{-\gamma}$, where the scaling exponent $\gamma$ is typically in the range $2<\gamma<3$. These networks are scale-free, greatly vary in size and structural complexity, but similar in that most nodes have just a few connections, and some have a vast amount of links. For instance, in the cellular metabolic network, most molecules participate in just one or two biochemical reactions, and some molecules, such as water and adenosine triphosphate are discovered in most reactions. It forms a striking contrast to random networks that follow a bell-shaped Poisson degree distribution $P_k=e^{-\langle k\rangle} \langle k\rangle^{-k}/k!$, where most nodes have approximately the same number of links.

Our ultimate goal is to design or employing existing machine learning methods to train voting rules with certain axiomatic properties. The learnability for voting rules satisfying some desirable fairness axioms is very useful. Once a new voting rule is proposed, it may have some special fairness properties. It’s pretty hard to handcraft another voting rule with the same axiom. However, we could learn a similar voting rule from instances, or even a simpler rule with the exactly same property. It’s promising to learn all those existing voting rules efficiently and have their good genes to design a single meta-voting rule, that satisfies more fairness criteria, or at least with a higher satisfiability to all fairness criteria than any existing voting rule.