Pruning-based analysis of input sensory relevance in decentralized controllers of voxel-based soft robots

This repository contains the code valid for the Bio-Inspired Artificial Intelligence course project of the master Artificial Intelligence Systems at the University of Trento, Italy.

The implementation of the hnn module has taken inspiration from the SBM repository.

The code simulates the evolution of a population of soft robots controlled by decentralized controllers, one for each voxel. The controllers are neural networks with a fixed topology, and the weights are updated through Hebbian learning. The cma-es evolutionary algorithm is used to optimize the ABCD rules of the Hebbian learning.

The environment is simulated using the evogym library.

Getting started

Installation

The code supports Python 3.8 and above. To install the required dependencies, first it is required to install the evogym library. To do so, follow the instructions in the official repository.

Then, install the required dependencies using pip:

pip install -r requirements.txt

Running the code

The code can be run using the main.py script in the src directory. The script accepts the following arguments:

Parameter	Description	Required	Default
--robot	The robot used. At the moment, the qorm is the only supported.	False	worm
--env	The environment of the evogym library to simulate the robot. Supported values are walking_flat, down_stepper and _soft_bridge	False	walking_flat
--network	The architecture of the networks assigned to the controllers. Supported values are hnn	False	hnn
----evo-algo-type	The evolutionary algorithms. Supported values are cma-es	False	hnn
--nodes	The nodes of the architecture chosen. The considered inputs are 15 and the outputs are 2. It can accepts also hidden layers.	False	[15, 2]
--eta	The eta of the nueral networks.	False	0.1
--robot-structure-path	The path where to store/load the structure of the robot.	False	../data/robot_structure/worm/default.json
--random-structure	Whether to generate a random structure for the robot or not.	False	False
--train	Whether to train the robot.	False	False
--test	Whether to test the robot.	False	False
--prune	Whether to prune the robot.	False	False
--weight-path	The path to store/load the controllers ABCD rules.	True
--generations	The number of generations to train the robot.	False	30
--offsprings	The number of offsprings every generation defines.	False	15
--population-size	The size of the population of the evolutionary algorithm.	False	4
--sigma	The sigma value of the evolutionary algorithm.	False	4
--max_steps	The maximum steps the individual do in the environment.	False	2000
--weight_update_steps	The number of steps the individual do before updating its weights.	False	150
--prune_ratio	The prune ratio to apply during the pruning phase.	False	60
--weight_pruning_time	The update weights time where to apply the pruning.	False	5
--multi-processing	Supported for training, it allows to run multiple individuals simulations in parallel.	False	False
--raise-error-in-case-of-loading-structure-path	Whether to raise error in case the robot structure path is wrong or used the default structure.	False	True
--display	Whether to display the robot movements during test and pruning simulations.	False	False