Skip to content

Latest commit

 

History

History
98 lines (79 loc) · 5.61 KB

README.md

File metadata and controls

98 lines (79 loc) · 5.61 KB
Raw

BipartiteRegulatorProbing

We get a complete bipartite graph $G = (A \cup B, A \times B)$ with a set $A$ of $\mathit{Regulators}$ and a set $B$ of $\mathit{Positions}$. Every edge $(a,b) \in A \times B$ has an independent discrete distribution $D_{a,b}$ over the same support $\mathcal{V} :=$ { $0,...,|\mathcal{V}| - 1$ }. We know $D_{a,b}$ but not their edge weight realizations $w_{a,b} \sim D_{a,b}$. We can now $\mathit{probe}$ $k$ $\mathit{Regulators}$ thus revealing their incident edge weights. At the end, we have to choose $\ell$ $\mathit{Regulators}$ among the $\mathit{probed}$ ones to maximize a given set goal function $f$.

Usage

The algorithms are implemented in Rust. Therefore, Rust must be installed in advance. On most Unix-like systems, you can install Rust with:

curl --proto '=https' --t1sv1.2 -sSf https://sh.rustup.rs | sh

After that you can modify the code as you wish and compile using the preinstalled cargo package manager:

cargo build --release

Input Format

The input format is strictly dictated and cannot be changed without changing the code. You can create Instances of BPR using Python3 and scripts/createData.py with the following command:

python3 scripts/createData.py --number <Number of Graphs> \
    --na_min <Minimum Number of Regulators> \
    --na_max <Maximum Number of Regulators> \
    --nb_min <Minimum Number of Positions> \
    --nb_max <Maximum Number of Positions> \
    --vs_min <Minimum Size of Support> \
    --vs_max <Maximum Size of Support> \
    --output <Output Directory> 
    [--name <Custom Name>]

You can see an example file here: EXAMPLE

Alternatively, you can create graph instances at runtime with prespecified parameters

Running the algorithms

As there are multiple possibilities of choosing $k$ and $\ell$ for every graph with $n_A \mathit{Regulators}$, there are multiple ways to run algorithms all using the same main binary created. The standard prefix is always

target/release/bpr --file <Input Graph file> \
    --log <Path to log-file> \
    --instances <Number of Instances> \
    --goal <Goal Function> \
    --algorithm <Algorithm> \
    --parameters <Number> \
    [--not-opt]

Parameters is used to run on every possible fraction-combination of $k$ and $\ell$. Namely, if parameters $= 2$, then the algorithms will run on $k = \frac{1}{3}n_A, \frac{2}{3}n_A$ and $\ell = \frac{1}{3}k, \frac{2}{3}k, k$.

If you wish to create graph instances at runtime, instead use

target/release/bpr --log <Path to log-file> \
    --na <Number of Regulators> \
    --nb <Number of Positions> \
    --vs <Size of Support> \
    --iterations <Number of Graph Instances> \
    --instances <Number of Instances per Graph Instance> \
    --parameters <Parameters as above> \
    --goal <Goal Function> \
    --algorithm <Algorithm> 
    [--poisson] <Should Support model a Poisson Distribution>
    [--not-opt]

Algorithms

Goal Input Name Runtime Approximation Factor Source
MAX / SUM OPT OptimalOfflineAlgorithm $\mathcal{O}(n_A \cdot \log n_A)$ $OPT$ -
MAX / SUM AMP AdaptiveMyopicPolicy $\mathcal{O}(n_A \cdot (k + \log n_A))$ $\frac{e - 1}{e}OPT_A$ SMSM
MAX / SUM NAMP NonAdaptiveMyopicPolicy $\mathcal{O}(n_A \cdot \log n_A)$ $\frac{e - 1}{2e}OPT_A$ SMSM
COV OPT OptimalOfflineAlgorithm $\mathcal{O}(\ell \cdot n_A \cdot n_B)$ $\frac{e - 1}{e}OPT$ MSM
COV AMP AdaptiveMyopicPolicy $\mathcal{O}(k^2 \cdot \ell \cdot n_A \cdot n_B)$ - SMSM
COV NAMP NonAdaptiveMyopicPolicy $\mathcal{O}(k^2 \cdot \ell \cdot n_A \cdot n_B)$ - SMSM

To run all algorithms on a specified goal, use ALL and use --not-opt if you do not want to run OPT - otherwise it is always run and logged.

Goal Functions

There are $3$ possible goal functions. $f_{max}, f_{sum}$ which both reduce to Top-l-ProbeMax and $f_{cov}$ which reduces to a variation of MaximumCoverage.

For $f_{max}$ and $f_{sum}$, each $\mathit{Regulator }$ $a \in A$ is assigned an independent value, namely the maximum or the sum of all its incident edges. After that, we have to choose $\ell$ $\mathit{Regulators}$ to maximize the sum of their values.

For $f_{cov}$, each $\mathit{Position}$ $b \in B$ is assigned the value of the highest incident edge to a $\mathit{Regulator}$ $a \in S$ in the chosen probed subset $S \subseteq A$. We have to choose $\ell$ probed $\mathit{Regulators}$ to maximize the sum of all $\mathit{Position}$-values.

Jobs

The jobs folder contains all bash files to run the algorithms for comparison on the Goethe-HHLR cluster.

Scripts

The scripts folder contains all script files. Note that Python must be installed beforehand. The scripts are:

  • Data Generation using createData.py
  • Data Compression using bprTimeCompression.py, bprParametersCompression.py or bprValuesCompression to compress logged results into small data that then can be plotted
  • Plotting Data using timePlot.py, parametersPlot.py or valuesPlot.py

Installing the necessary python packages can be done via

pip install -r scripts/requirements