This model visualizes gradient descent optimization - the fundamental algorithm used to train neural networks and other machine learning models. Agents represent different optimization algorithms searching for the minimum of a loss landscape (the “error surface” that ML models try to minimize during training).

The model demonstrates how different optimizer types (SGD, Momentum with different parameters) behave on various loss landscapes, from simple bowls to the notoriously difficult Rosenbrock “banana valley” function. This helps build intuition about why certain optimization algorithms work better than others for different problem geometries.

HOW IT WORKS

…

An agent-based model which explores Creativity and Urban Development

The ForagerNet3_Demography model is a non-spatial ABM designed to serve as a platform for exploring several aspects of hunter-gatherer demography.

Zombies move toward humans and humans move (faster) away from zombies. They fight if they meet, and humans who lose become zombies.

The purpose of the model is to simulate the future growth of human settlements in the Nile river valley in Egypt. The model contains processes to mimic spatial patterns found in the case study region.

BarterNet is a platform for modeling early barter networks with the aim of learning how supply and demand for a good determine if traders will learn to use that good as a form of money. Traders use a good as money when they offer to trade for it even if they can’t consume it, but believe that they can subsequently trade it for a good they can consume in the near future.

The simulation generates two kinds of agents, whose proposals are generated accordingly to their selfish or selfless behaviour. Then, agents compete in order to increase their portfolio playing the ultimatum game with a random-stranger matching.

An empirical-response agent-based model of how many personalized feeds execute a shared low-exposure creator-discovery objective. Built from the KuaiRec dataset: the big interaction matrix initializes a transparent rank-8 matrix-factorization platform learner and the activity schedule, while the near-complete small matrix returns observed viewing responses only after a user-video pair is exposed. Four exploration policies (synchronous low-exposure targeting, uniform exploration, per-user random tie-breaking, capacity-balanced coordination) are compared over 28 rounds at a nominal 10% exploration budget, across 30 paired seeds (core) and 10 paired seeds (bias-only probe), with slot-level redundancy, cross-user collision, and coverage diagnostics.
This release accompanies an anonymised manuscript under review at the Journal of Artificial Societies and Social Simulation.

The model represents 1,411 users, 3,327 videos, 2,031 authors, and an adaptive platform over 28 discrete rounds derived from the KuaiRec big-matrix activity calendar. Exploration policies differ only in how a fixed 10% slot budget is allocated; all policies share the opportunity schedule, response oracle, initial checkpoints, and online update rule.

Archive contents: analysis pipeline scripts (01-16), frozen machine-readable protocols with input hashes, initial model checkpoints, aggregate result tables, the complete ODD protocol record, and publication figures. Raw KuaiRec files are not redistributed; obtain them from the official dataset repository and verify against the input hashes in data_contract/. Row-level oracle tables are excluded by design.
…

Status-power dynamics on a playground, resulting in a status landscape with a gender status gap. Causal: individual (beauty, kindness, power), binary (rough-and-tumble; has-been-nice) or prior popularity (status). Cultural: acceptability of fighting.

An agent-based framework that aggregates social network-level individual interactions to run targeting and rewarding programs for a freemium social app. Git source code in https://bitbucket.org/mchserrano/socialdynamicsfreemiumapps