B.BALARAM B.FULL-STACK DEVELOPER

HOME/WORK/SEMAIS

MULTI-AGENT AI / EVOLUTIONARY SIMULATION

SEMAIS

A self-evolving simulation engine where a population of agents learns to coordinate under constraints, changing environments and adversarial pressure.

ROLE
Sole developer
CONTEXT
Independent project
STACK
Python
FOCUS
Multi-agent AI

01THE QUESTION

How can multiple agents coordinate under constraints, changing environments and adversarial pressure?

Hand-written rules for each agent do not scale to that question. Behaviour that works in one environment fails when the environment shifts or an opponent adapts. SEMAIS asks a different question: instead of designing the behaviour, can a population discover it?

02THE APPROACH

I built a self-evolving multi-agent simulation engine in Python. Each agent carries a 7-parameter behavioural genome. A genetic algorithm evolves that genome over 50+ generations across 100+ agents.

Agents navigate with A* pathfinding and talk to each other over fault-tolerant communication, so one failed link does not stall the group. Afterwards, unsupervised clustering groups the evolved genomes into 4–5 distinct strategies.

7PARAMETER BEHAVIOURAL GENOME
50+GENERATIONS OF EVOLUTION
100+AGENTS IN THE POPULATION
4–5EMERGENT STRATEGIES FOUND

03HOW IT
WORKS

A simplified view of the flow, step by step.

  1. Spawn

    A population of 100+ agents, each with its own 7-parameter genome.

  2. Act

    Agents path-find with A* and exchange messages over fault-tolerant links.

  3. Select

    The genetic algorithm keeps the genomes that performed best.

  4. Evolve

    Selected genomes are recombined and mutated into the next generation.

LOOPRepeat for 50+ generations.
THENCluster the final genomes into 4–5 emergent strategies.

ILLUSTRATIVE DIAGRAM · BUILT FOR THIS SITE · NOT THE ORIGINAL SYSTEM

04TRY
THE IDEA

This is a small illustrative simulation of the ideas above: agents moving and adapting on a field. It is not the original Python engine. Move your pointer or tap to disturb it.

05BUILT
WITH

PYTHONGENETIC ALGORITHMSMULTI-AGENT SYSTEMSA* PATHFINDINGCLUSTERING

THE DIAGRAM AND SIMULATION ARE CONCEPTUAL ILLUSTRATIONS BUILT FOR THIS SITE. THE FIGURES (7-PARAMETER GENOME, 50+ GENERATIONS, 100+ AGENTS, 4–5 STRATEGIES) COME FROM MY RESUME.

Want something like this built?