Bayesian neuroevolution using distributed swarm optimisation and tempered MCMC

A synergy of Neuro-evolution and Bayesian neural learning by using Particle Swarm Optimization for forming efficient proposals in parallel tempering MCMC. The architecture is implemented in a parallel computing environment for improving computational efficiency.

Requirements

This project was tested on python 3.8 and the required packages can be installed using the provided requirements.txt:

    pip install -r requirements.txt

Running Evolutionary Parallel Tempering

    cd pt_pso
    python pt_pso_mcmc_pop.py --problem <problem-name>

The script supports following additional command line arguments:

usage: evo_pt_mcmc.py [-h] [--problem PROBLEM] [--num-chains NUM_CHAINS]
                      [--population-size POPULATION_SIZE]
                      [--num-samples NUM_SAMPLES]
                      [--swap-interval SWAP_INTERVAL] [--burn-in BURN_IN]
                      [--max-temp MAX_TEMP] [--topology TOPOLOGY]
                      [--run-id RUN_ID] [--root ROOT]
                      [--train-data TRAIN_DATA] [--test-data TEST_DATA]
                      [--config-file CONFIG_FILE]

Run Evolutionary Parallel Tempering

optional arguments:
  -h, --help            show this help message and exit
  --problem PROBLEM     Problem to be used for Evolutionary PT: "synthetic",
                        "iris", "ions", "cancer", "bank", "penDigit", "chess", "TicTacToe"
  --num-chains NUM_CHAINS
                        Number of Chains for Parallel Tempering
  --population-size POPULATION_SIZE
                        Population size for G3PCX Evolutionary algorithm.
  --num-samples NUM_SAMPLES
                        Total number of samples (all chains combined).
  --swap-interval SWAP_INTERVAL
                        Number of samples between each swap.
  --burn-in BURN_IN     Ratio of samples to be discarded as burn-in samples.
                        Value 0.1 means 10 percent samples are discarded
  --max-temp MAX_TEMP   Temperature to be assigned to the chain with maximum
                        temperature.
  --topology TOPOLOGY   String representation of network topology. Eg:-
                        "input,hidden,output"
  --run-id RUN_ID       Unique Id to identify run.
  --root ROOT           path to root directory (evolutionary-pt).
  --train-data TRAIN_DATA
                        Path to the train data
  --test-data TEST_DATA
                        Path to the test data
  --config-file CONFIG_FILE
                        Path to data config yaml file

Note: The default values of these values can be changed from config.py and data.yaml file

DataSets - Classification

The approach is tested on the following datasets:

1. Iris
2. Ionosphere
3. Breast Cancer
4. Pen Digit

Cite this work

@article{kapoor2022bayesian,
  title={Bayesian neuroevolution using distributed swarm optimization and tempered MCMC},
  author={Kapoor, Arpit and Nukala, Eshwar and Chandra, Rohitash},
  journal={Applied Soft Computing},
  volume={129},
  pages={109528},
  year={2022},
  publisher={Elsevier}
}