Predicting the matching probability and the expected ride/shared distance for each dynamic ridepooling order in large scale (grid) network

Prediction model

Seekers and takers

Prediction model

For each seeker-state:

• $p_{s(w)}$: probability of a seeker getting matched in this state.

For each taker-state:

• $p_{t(a, w)}$: probability of a taker getting matched in this state.
• $\lambda_{t(a, w)}$: arrival rate of unmatched passengers into this state.
• $\rho_{t(a, w)}$: probability of having at least one taker in this state.
• $\eta_{t(a, w)}$: aggregate arrival rate of matching opportunities for takers in this state.

For each match pair:

• $\eta_{t(a, w)}^{s(w)}$: mean arrival rate of matching opportunities in the seeker-state for takers in the taker-state.

There are complex interactions among these variables, so we formulated the interactions into a fixed point problem: $$ F(\mathbf{X}) = \mathbf{X}, \mathbf{X} = (\mathbf{\lambda}, \mathbf{p_s}, \mathbf{p_t}, \mathbf{\rho}, \mathbf{\eta})^T $$

Pre-Experiments

Parameters:

• Pick-up time: 1s
• Maximum detour distance: 5m
• Search Radius: 5m
• Speed: 1m/s
• Demand rate of an OD: 0.5/s
• Simple fixed-point iteration method with stopping criteria $\Vert \mathbf{X}^{(k+1)} - \mathbf{X}^{(k)}\Vert_\infty \leq 1%$

$$ \begin{matrix} \rm Network & 3030 & 180180 & 180*180 \ \hline \rm OD & 300 & 2000 & 1e4 \ \hline \rm Taker & 5056 & 191956 & 9.5e5 \ \hline \rm Match & 32772 & 148530 & 3.6e6 \ \hline \rm Iteration & 29 & 20 & 25 \ \hline \rm Step 0 / s & 0.488 & 108.342 & 2692.946 \ \hline \rm Step 1 / s & 0.453 & 4.285 & 53.341 \ \hline \end{matrix} $$

Clustering

Preliminaries

The input of spectral clustering is a Similarity Matrix $S\in \mathbb{S}^n$ and number of clusters $k$. Every element of the similarity matrix $w_{ij}$ measures the pairwise similarity between node $i$ and $j$ on a graph $G(V, E)$. Our objective is to minimize RatioCut or Ncut: $$ \mathrm{RatioCut}(A_1, \cdots, A_k) = \frac{1}{2}\sum\limits_{i=1}^k \frac{W(A_i, \overline{A_i})}{|A_i|}\ \mathrm{Ncut}(A_1, \cdots, A_k) = \frac{1}{2}\sum\limits_{i=1}^k \frac{W(A_i, \overline{A_i})}{\mathrm{vol}(A_i)} $$

where $A_i$ denotes a set of nodes in the $i$th cluster, $\overline{A_i}$ denotes those are not in the cluster, $|A_i|$ denotes number of nodes in the $i$th cluser and: $$ \mathrm{vol}(A_i) = \sum\limits_{k \in A_i}\sum\limits_{j = 1}^n w_{kj}\ W(A, B) = \sum\limits_{i\in A, j\in B}w_{ij} $$

Minimize RatioCut approximately

Define $H\in \mathbb{R}^{n\times k}$ as an indicator matrix which indicates whether node $i$ belongs to cluster $j$. That is: $$ h_{ij} = \begin{cases} \frac{1}{\sqrt{|A_j|}}, v_i \in A_j\ 0, \mathrm{otherwise} \end{cases} $$ Then we can get that: $$ \mathrm{RatioCut}(A_1, \cdots, A_k) = \mathbf{Tr}(H^TLH) $$ where $L$ is the Laplacian Matrix of the graph: $$ L = D - W = \mathbf{diag}(\sum\limits_{j=1}^n w_{1j}, \cdots, \sum\limits_{j=1}^n w_{nj}) - (w_{ij}){n\times n} $$ So the optimization problem is: $$ \min \mathbf{Tr}(H^TLH)\ \mathrm{s.t.\ } H = (h{ij}){n\times k}, h{ij} = \begin{cases} \frac{1}{\sqrt{|A_j|}}, v_i \in A_j\ 0, \mathrm{otherwise} \end{cases} $$

This integer programming problem is a NP hard problem. But we can relax it by allowing $h_{ij}$ to take any real values. Then the relaxed problem becomes: $$ \min \mathbf{Tr}(H^TLH)\ \mathrm{s.t.\ } H^TH = I $$ where $H^TH = I$ can be checked under our definition of $H$.

It is a standard Trace Minimization Problem and according to Rayleigh-Ritz Theorem the solution is given by choosing the first $k$ eigenvectors of $L$ as columns of $H$.

Finally, we should re-convert the real values of $H$ into discrete partition. Here we can use k-means algorithms on the rows of $H$.

Experiments

We use spectral clustering to assign 10000 ODs on 360*360 grid network into 5 clusters and each cluster has 2993, 624, 748, 2801, 2822 ODs respectively. The runtime for global variable-generation is 5549s(92.5min) while the one for each cluster is 540, 31, 45, 459, 370s(24.1min totally) respectively. The runtime for global iteration is 68s while the one for each cluster is 17, 3, 4, 13, 14s(51s totally) respectively when the absolute error $\epsilon = 10^{-4}$.

Violin plot