An evolutionary multi-agent platform based on mesa and NEAT

Overview

EvoAgent

An evolutionary multi-agent platform based on mesa and NEAT
A video of an old version of EvoAgent can be found on youtube:
https://www.youtube.com/watch?v=pOF1H84xPik&ab_channel=VBiscione

Overview

With this tool you can evolve agents in a simple 2D environments.
At the beginning, agents are endowed with a simple neural network which is specified by their genome. With time, they'll make children, which will inherit the genome, plus some random mutations. The mutations will add a connetion, a node, or change the weight. The evolutionary system is based on NEAT by Kenneth Stanley (paper here). However, the main difference between this approach and the classic evolutionary algorithm approach, is the naturalistic approach we are using here. In a classic evolutionary algorithm we would use a fitness function, we would test our agents, and we would use a genetic operator for generating the next generation. We have none of that here. Agents will spawn a new children every x time-steps. The more they survive, the more children they'll get. I found this approach much more intuitively valid as will simply respond to the rule: the more you survive, the more you'll spread your genes. This will also allow for generational interactions (as multiple generations will live at the same time), in case one wants to run experiments on family behaviour etc.

Install

git clone [email protected]:ValerioB88/EvoAgent.git

EvoAgent plots everything in the browser. This is heavily dependent on the Mesa library. My fork of Project Mesa should be automatically added as a submodule.

Examples

There are several examples in the folder code/evoagent/experiments/, which should get you started. Most of this is self explanatory. For example, run the file difficult_env.py and a browser with the environment view should open. Both the model state and the population gets saved (the model state includes the population). You want to use the model state to continue a running simulation with exactly the same parameters, from exactly the same point (the model state includes the population state). Instead, you may want to test an evolved population on a different enviornment: in that case, load the population. Examples of both these cases in continue_sim.py and load_pop.py. Model and population get saved every 1000 time steps, or when you press s.

Environment View

This contains a canvas, information about the selected agent, and some plots. At each points there always be a selected agents, for which many info are shown in the text field on the right sidebar. You can change the selected agent by pressing a and d on your keyboard (in a future version, it will be possible to click on the agent on the canvas to select it). Other info are shown in the following image:

Shortcuts

q start/stop simulation
a/d previous/next agent
k kill selected agent
c selected agent spawn a child
s save model
x advance the simulation one step

Agents Behaviour

Agents will have a limited field of view, specified by the parameters fov and max_vision_dist. You can visualize them by checking Render FOV in the environment view. Each agent will have a maximum lifespan of max_age = 500, but it can die faster if its energy goes to 0. They have a energy_depletion rate of 0.008. An agent will eat a food token when it touches it, getting an energy increase of 0.8 (or whathever established by the current Epoch, see later), up to 1. Agent will become fertile after fertile_age_start = 90 timesteps. At that point, they will spawn a child every time_between_children = 100 timesteps. For now, reproduction is asexual, so there is no crossover (this will be expanded in a next version).

Agents can only perceive the closest food token, IF it's within their field of view. If the selected agent is perceiving a food token, this will be represented with a red color in the environment view. Agents' network input will be the linear distance to the perceived food token, the angular distance, and their own energy level. There will be two output units, one defining the forward/backward movement, the other one defining the rotation.

Epochs

The epochs are interval of times defined by the amount of food in the environment.

epochs = deque([Epoch(2000, [0.8]*200),
                Epoch(2000, [0.8]*100),
                Epoch(6000, [0.8]*75)])

This means that we have 3 epochs, the first lasting 2000 iteration, the last 6000 iterations. The first will have 200 food tokens, each provigin 0.8 units of energy (recall that [0.8]*200 generates a list of 200 elements of value 0.8: [0.8, 0.8, 0.8....]. By changing the second parametr, you can specify the distribution of food token easily. You can also have "posionous" food: Epoch(2000, [-0.2]*200) this will work. However, currently the agents don't have a way to differentiate between "good" and "bad" food, so it doesn't make much sense to use bad food right now.

The simulation will activate one epoch at the time, starting from the first one. When the last epoch finishes, the simulation stop. You can instead run the simulation indefinitively by setting the epoch duration to np.inf.

Speed Up Simulation

In the environment view, uncheck Render Canvas, then set the Frame per Second slider to 0. That's as fast as it can get!

Known Problems

If you stop the simulation and try to run it again, you'll get the error:
OSError: [WinError 10048] Only one usage of each socket address (protocol/network address/port) is normally permitted
To fix this, try to change the port in server.launch(...). If it doesn't work, close and open your browser. I know this is annoying, but hadn't have the time to fix it yet.

This has only been tested on Windows 10 and Edge 97. I don't plan to test it on any other configuration. Sorry!

To Do

☐ Add sexual reproduction: find a partner, do crossover
☐ Many agents options should be in a config file
☐ Select an agent by clicking on it with the mouse
☐ Implement data collection - or adapt mesa data collection
☐ Implement more ways to perceive food
☐ Implement ways to perceive other agents
☐ Server port stays open after code stops - which often lead to annoying behaviour. Find a way to fix this

You might also like...
An unofficial python API for trading on the DeGiro platform, with the ability to get real time data and historical data.

DegiroAPI An unofficial API for the trading platform Degiro written in Python with the ability to get real time data and historical data for products.

A platform for developers 👩‍💻  who wants to share their programs and projects.
A platform for developers 👩‍💻 who wants to share their programs and projects.

Fest-Practice-2021 This project is excluded from Hacktoberfest 2021. Please use this as a testing repo/project. A platform for developers 👩‍💻 who wa

This repository provides a set of easy to understand and tested Python samples for using Acronis Cyber Platform API.

Base Acronis Cyber Platform API operations with Python !!! info Copyright © 2019-2021 Acronis International GmbH. This is distributed under MIT licens

Tools, guides, and resources for blockchain analysts to interface with data on the Ergo platform.

Ergo Intelligence Objective Provide a suite of easy-to-use toolkits, guides, and resources for blockchain analysts and data scientists to quickly unde

Coinloggr - A learning resource and social platform for the coin collecting community
Coinloggr - A learning resource and social platform for the coin collecting community

Coinloggr A learning resource and social platform for the coin collecting commun

Program to send ROM files to Turbo Everdrive; reverse-engineered and designed to be platform-independent
Program to send ROM files to Turbo Everdrive; reverse-engineered and designed to be platform-independent

PCE_TurboEverdrive_USB What is this "TurboEverdrive USB" thing ? For those who have a TurboEverdrive v2.x from krikzz.com, there was originally an opt

Fully cross-platform toolkit (and library!) for MachO+Obj-C editing/analysis
Fully cross-platform toolkit (and library!) for MachO+Obj-C editing/analysis

fully cross-platform toolkit (and library!) for MachO+Obj-C editing/analysis. Includes a cli kit, a curses GUI, ObjC header dumping, and much more.

An ultra fast cross-platform multiple screenshots module in pure Python using ctypes.

Python MSS from mss import mss # The simplest use, save a screen shot of the 1st monitor with mss() as sct: sct.shot() An ultra fast cross-platfo

The Open edX platform, the software that powers edX!

This is the core repository of the Open edX software. It includes the LMS (student-facing, delivering courseware), and Studio (course authoring) compo

Owner
Valerio1988
Valerio1988
InfiniPy has some neat features - like the endpoint for function

InfiniPy has some neat features - like the endpoint for function

ZeroTwo 7 Nov 20, 2022
PressurePlate is a multi-agent environment that requires agents to cooperate during the traversal of a gridworld.

PressurePlate is a multi-agent environment that requires agents to cooperate during the traversal of a gridworld. The grid is partitioned into several rooms, and each room contains a plate and a closed doorway.

Autonomous Agents Research Group (University of Edinburgh) 6 Dec 3, 2022
A multi-platform fuzzer for poking at userland binaries and servers

litefuzz A multi-platform fuzzer for poking at userland binaries and servers litefuzz intro why how it works what it does what it doesn't do support p

null 52 Nov 18, 2022
Speed up your typing by some exercises in the multi-platform(Windows/Ubuntu).

Introduction This project purpose is speed up your typing by some exercises in the multi-platform(Windows/Ubuntu). Build Environment Software Environm

lyfer233 1 Mar 24, 2022
The learning agent learns firstly approaching to the football and then kicking the football to the target position

Football Court This project utilized Pytorch and Tensorflow so that the learning agent learns firstly approaching to the football and then kicking the

null 1 Nov 19, 2021
A software dedicated to automaticaly select the agent of your desire in Valorant

AUTOPICKER A software dedicated to automaticaly select the agent of your desire in Valorant GUIDE Before stariting to use this program check if you ha

p1n00 0 Sep 24, 2022
The Python agent for Apache SkyWalking

SkyWalking Python Agent SkyWalking-Python: The Python Agent for Apache SkyWalking, which provides the native tracing abilities for Python project. Sky

The Apache Software Foundation 149 Dec 12, 2022
Automatic and platform-independent unpacker for Windows binaries based on emulation

_ _ __ _ __ _ | | | | / / (_) \ \ | | | | | |_ __ | | _ | | _ __ __ _ ___| | _____ _ __

null 514 Dec 21, 2022
Direct Multi-view Multi-person 3D Human Pose Estimation

Implementation of NeurIPS-2021 paper: Direct Multi-view Multi-person 3D Human Pose Estimation [paper] [video-YouTube, video-Bilibili] [slides] This is

Sea AI Lab 253 Jan 5, 2023
Cross-platform config and manager for click console utilities.

climan Help the project financially: Donate: https://smartlegion.github.io/donate/ Yandex Money: https://yoomoney.ru/to/4100115206129186 PayPal: https

null 3 Aug 31, 2021