Skip to content

README.md

Latest commit

 

History

History
88 lines (63 loc) · 2.23 KB

README.md

File metadata and controls

88 lines (63 loc) · 2.23 KB

Benchmarks

This folder holds code to reproduce the results from our paper

@article{lineflow,
  author = {Kai M{\"u}ller and Martin Wenzel and Tobias Windisch},
  title = {LineFlow: A framework to learn active control of production lines},
  year = {2025},
}

Scenarios

Part distribution

Run part_distribution/run_part_distribution_k.sh with k=3,4,5.

Waiting time

Run waiting_time/run_waiting_time.sh for WT and waiting_time/run_waiting_time_jump.sh for WTJ

Worker assignment

Run worker_assignment/run_worker_assignmen_k.sh with k=3,4,5.

Complex line

Run complex_line/run_complex_line.sh to run all complex line benchmarks, i.e., recurrent PPO with and without curriculum learning as well as stacked PPO with curriculum learning.

Parallelization

The central file is train.py, which is triggered with specific parameters for each benchmark scenario and RL algorithm. More specifically, each benchmark scenario triggers another run script for each algorithm, for instance run_a2c.sh which looks like:

#!/bin/bash

SEED=$1
ENV=$2
N=$3
INFO=$4
TOTAL_STEPS=$5

LR=("0.01" "0.001" "0.0005" "0.01" "0.001" "0.0005")
ENTCOEF=("0.0"  "0.0"   "0.0"    "0.001" "0.001" "0.001")

for i in "${!LR[@]}"; do
    echo "Running config $i: lr=${LR[$i]}, ent_coef=${ENTCOEF[$i]}"
    python3 train.py --deterministic \
        --env="$ENV" --n_cells="$N" --model=A2C \
        --seed="$SEED" --info="$INFO" --total_step="$TOTAL_STEPS" \
        --learning_rate="${LR[$i]}" \
        --ent_coef="${ENTCOEF[$i]}"
done

In our experiments, the benchmarks where executed within a slurm cluster and these shell scripts looked like:

#!/bin/bash
#SBATCH --partition=<PARTITION>
#SBATCH --mem=10G                 
#SBATCH --ntasks=1
#SBATCH --cpus-per-task=5
#SBATCH --array=0-5%3             

SEED=$1
ENV=$2
N=$3
INFO=$4
TOTAL_STEPS=$5

declare -a LR=("0.01" "0.001" "0.0005" "0.01" "0.001" "0.0005")
declare -a ENTCOEF=("0.0"  "0.0"   "0.0"    "0.001" "0.001" "0.001" )

srun python3 train.py --deterministic \
    --env=$ENV --n_cells=$N --model=A2C \
    --seed=$SEED --info="$INFO" --total_step=$TOTAL_STEPS \
    --learning_rate="${LR[$SLURM_ARRAY_TASK_ID]}" \
    --ent_coef="${ENTCOEF[$SLURM_ARRAY_TASK_ID]}"