Network optimizer with genetic algorithm

This repository is in response to Bitwise code challenge 2022 2/3. There is a network of nodes that are connected with some edges. The goal is to remove as much edges as possible, while maximizing total score P on

$$ P = (P_{og} / P_r - 1) \cdot 1000 \,,$$

where $P_{og}$ is the original score, $P_r$ new score of a network, where edges are removed, and $P_i$ are calculated by

$$ P_i = A Wt + B D{avg} \, . $$

Here, $A=0.1$, $B=2.1$, $Wt$ is the sum of length of all arcs and $D{avg}$ average shortest path length from each individual point to any other point.

Original data is included in this repository at bitwise_challenge_2022_2/koodipahkina-data.json.

Quickstart

Requirements:

• Python 3.10
• Poetry (see installation instructions)

poetry install  # Install environment and dependencies
poetry shell    # Enter environment
python -m bitwise_challenge_2022_2  # Run optimization

Please see program help for options:

$ python -m bitwise_challenge_2022_2 --help
usage: __main__.py [-h] [--n-trials N_TRIALS] [--optuna-prune]
                   [--restart RESTART] [--metric-log METRIC_LOG]
                   [--xpath XPATH] [--resume] [--multiprocessing] [--quiet]
                   [--plot] [--seed SEED] [--elite-frac ELITE_FRAC]
                   [--mutant-frac MUTANT_FRAC] [--elite-bias ELITE_BIAS]

Optimize liana network.

options:
  -h, --help            show this help message and exit
  --n-trials N_TRIALS   Number of trials for hyperparameter optimization.
                        Default: None
  --optuna-prune        Use pruning in hyperparameter optimization. Default:
                        no pruning
  --restart RESTART     Restart specification in format
                        key1:value1,key2:value2. Available values: see keyword
                        arguments at
                        https://pymoo.org/interface/termination.html .
                        Default: n_max_evals:1000000,n_last:500,n_max_gen:1000
  --metric-log METRIC_LOG
                        Location for logging optimization running metrics.
                        Default: opt_log.txt
  --xpath XPATH         Location for saving intermediate x values for
                        population. Default: opt_X_latest.npy
  --resume              Resume from xfile
  --quiet, -q           Suppress output
  --plot                Display progress plot during calculation
  --seed SEED           Random seed. Default: 1
  --elite-frac ELITE_FRAC
                        Elite fraction of population. Default: 0,2
  --mutant-frac MUTANT_FRAC
                        Mutant fraction of population. Default: 0.1
  --elite-bias ELITE_BIAS
                        Probability of elite gene transfer. Default: 0.7

Implementation

The optimization is built on pymoo optimization framework, more specifically its biased random key genetic algorithm implementation. A population of vectors is initialized, where vector length is equal to the number of original edges in the network. These vectors represent, whether the edge is to be removed ($x_i < 0.5$) or kept. Some initial vectors are constructed with a greedy algorithm, others generated randomly. All vectors are repaired so, that they produce a feasible solution:

1. find edge from set of not-connected edges that has minimum weight, and has one end in the graph and the other end in a not-connected node.
2. Add the edge to the graph.
3. Continue until graph is connected.

The genetic algorithm then is used to minimize the network score $P_r$.

A hyperparameter optimization framework optuna is also included, but tuning the BRKGA parameters did not result in any significant findings. For hyperparameter optimization, use flag --n-trials. By default, it is not used.

Numba Python compiler is used to make calculations faster. On an Intel i5-8350U CPU (Lenovo T480), the program runs at 600–800 evaluations per second, which results in a reasonable runtime.

For competition result, the optimizer was run on DataCrunch.io server with flag --restart n_max_evals:10000000,n_last:5000,n_max_gen:5000. A final score $P=1285.075$ was achieved. For the competition entry details, see also output from

git tag v1.4

Extras

If installed with poetry install --dev, a library for plotting is included. Thus intermediate logs by the optimizer can be visualized with scripts/plot_optimization_log.py; see help with

python scripts/plot_optimization_log.py --help