Artificial Immune Systems Package

A Python package for Artificial Immune Systems algorithms

---

Language

• Português

Documentation

• Official Docs
• Github Wiki

---

Introduction

AISP is a python package that implements artificial immune systems techniques, distributed under the GNU Lesser General Public License v3.0 (LGPLv3).

The package started in 2022 as a research package at the Federal Institute of Northern Minas Gerais - Salinas campus (IFNMG - Salinas).

Artificial Immune Systems (AIS) are inspired by the vertebrate immune system, creating metaphors that apply the ability to detect and catalog pathogens, among other features of this system.

What can you do with AISP?

AISP provides implementations of bio-inspired algorithms for:

• Anomaly detection: Identify abnormal patterns in data.
• Classification: Classify data with multiple classes.
• Optimization: Find optimal solutions for objective functions.
• Clustering: Group data without supervision.

---

Implemented Algorithms

Negative Selection (aisp.nsa)

• BNSA - Binary Negative Selection Algorithm
• RNSA - Real-Valued Negative Selection Algorithm

Clonal Selection (aisp.csa)

• AIRS - Artificial Immune Recognition System
• CLONALG - Clonal Selection Algorithm

Immune Network Theory (aisp.ina)

• AiNet - Artificial Immune Network for clustering and data compression

Module in Development

Danger Theory (aisp.dta)

• DCA - Dendritic Cell Algorithm (planned)

API overview

All algorithms follow a simple and consistent interface:

• fit(X, y, verbose: bool = True): trains the model for classification tasks.
• fit(X, verbose: bool = True): trains the model for clustering tasks.
• predict(X): makes predictions based on new data.
• optimize(max_iters: int =..., n_iter_no_change: int =..., verbose: bool = True): run the optimization algorithms

---

Installation

The module requires installation of python 3.10 or higher.

Dependencies

Packages	Version
numpy	≥ 1.22.4
scipy	≥ 1.8.1
numba	≥ 0.59.0

User installation

The simplest way to install AISP is using pip:

pip install aisp

---

Quick Start

Below are minimal examples demonstrating how to use AISP for different tasks.

Classification with RNSA

importnumpyasnpfromaisp.nsaimportRNSA# Generating training datanp.random.seed(1)
class_a=np.random.uniform(high=0.5, size=(50, 2))
class_b=np.random.uniform(low=0.51, size=(50, 2))
x_train=np.vstack((class_a, class_b))
y_train= ['a'] *50+ ['b'] *50# Training the modelmodel=RNSA(N=150, r=0.3, seed=1)
model.fit(x_train, y_train, verbose=False)
# Predictx_test= [
[0.15, 0.45], # Expected: 'a'
[0.85, 0.65], # Expected: 'b'
]
y_pred=model.predict(x_test)
print(y_pred)

Clustering with AiNet

importnumpyasnpfromaisp.inaimportAiNetnp.random.seed(1)
# Generating training dataa=np.random.uniform(high=0.4, size=(50, 2))
b=np.random.uniform(low=0.6, size=(50, 2))
x_train=np.vstack((a, b))
# Training the modelmodel=AiNet(
N=150,
mst_inconsistency_factor=1,
seed=1,
affinity_threshold=0.85,
suppression_threshold=0.7
)
model.fit(x_train, verbose=False)
# Predict cluster labelsx_test= [
[0.15, 0.45],
[0.85, 0.65],
]
y_pred=model.predict(x_test)
print(y_pred)

Optimization with CLONALG

importnumpyasnpfromaisp.csaimportClonalg# Define search spacebounds= {'low': -5.12, 'high': 5.12}
# Objective function (Rastrigin)defrastrigin(x):
x=np.clip(x, bounds['low'], bounds['high'])
return10*len(x) +np.sum(x**2-10*np.cos(2*np.pi*x))
# Initialize optimizermodel=Clonalg(problem_size=2, rate_hypermutation=0.5, bounds=bounds, seed=1)
model.register('affinity_function', rastrigin)
# Run optimizationpopulation=model.optimize(100, 50, False)
print(model.best_solution, model.best_cost) # Best solution

---

Examples

Explore the example notebooks available in the AIS-Package/aisp repository. These notebooks demonstrate how to utilize the package's functionalities in various scenarios, including applications of the RNSA, BNSA and AIRS algorithms on datasets such as Iris, Geyser, and Mushrooms.

You can run the notebooks directly in your browser without any local installation using Binder:

💡 Tip: Binder may take a few minutes to load the environment, especially on the first launch.