Example models¶

Henad has ten example models, in the set that henad::models::example_models() returns, and its own app and command line offer them. Your own project offers its own models, and can add some or all of the example models. Six of them run on the CPU, and four of those six have a GPU port running the same simulation entirely in compute shaders. The two network models run on the CPU only.

cargo run -p henad-cli -- --list

The four GPU entries appear only when wgpu finds an adapter with compute support. Each GPU port starts from its CPU counterpart's tick 0 bit for bit, which makes a comparison between the two backends fair.

Parameters¶

Every model declares its parameters, and both front ends read the same declarations. Print them with --params:

cargo run -p henad-cli -- boids --params

Each parameter is live or reload. A live parameter takes effect on the next tick, while a reload parameter applies only when the model is rebuilt. The engine prepends grid width and height to every grid model's list, and agent count, world width and world height to every agent model's list. A network model gets the same three parameters as an agent model, with num_agents as its node count and the world as the area in which its nodes are placed and drawn. A percentage parameter holds a fraction from 0 to 1, and the app shows it as a percentage.

Most models also declare actions, one-off changes to the state. The app draws each action as a button in the Parameters tab, and henad-cli --act runs an action at a given tick (see the command line). A GPU port declares the same actions as its CPU model, under the same ids.

Game of Life¶

Conway's Game of Life on a toroidal grid.

The randomise action refills the grid at the density that the model was built with, and clear kills every cell.

SIR Epidemic¶

The classic SIR compartmental model on a 2D grid with a Moore neighbourhood.

The seed_outbreak action infects each susceptible cell with the probability that initial_infected_pct had when the model was built.

Boids Flocking¶

Flocking over a torus. A spatial hash, rebuilt every tick, answers the neighbour queries.

The randomise_headings action points every boid in a random direction without changing its speed.

Ant Foraging¶

A population over a pheromone field, the one composite example model. Ants deposit into a scalar field that decays each tick, then steer by the values they read back.

The reset_colony action clears both pheromone trails and puts every ant back on the nest.

Virus on a Network¶

NetLogo's Virus on a Network, with rewiring and a directed variant added. Each infected neighbour of a susceptible node infects it with probability virus_spread_chance, independently of the others. Every virus_check_frequency ticks an infected node recovers with probability recovery_chance. A node that recovers becomes resistant with probability gain_resistance_chance, and susceptible again otherwise. Each node keeps its own check timer, started at a random offset. The stats count susceptible, infected and resistant nodes.

network picks the generator. Random draws a uniform random graph with average_node_degree * num_agents / 2 edges, rounded up. Geometric joins every pair of nodes closer than a radius chosen so that the mean degree comes out at average_node_degree. Either way each edge points in a random direction.

directed can be switched mid-run. Turned on, it lets the virus cross an edge only in that edge's direction. Turning it on or off keeps every edge and its direction.

keep_rewiring moves one random edge to a random unjoined pair of nodes each tick. The rewire action, the Rewire a link button, moves one edge the same way on demand. A rewire keeps the edge count and never joins a pair twice.

The model differs from the NetLogo version in these ways.

• NetLogo builds a spatially clustered network. It links a random node to its nearest unlinked node until the edge count is reached. Random keeps that edge count with no spatial structure, and Geometric is the spatial option.
• A run carries on after the last infected node recovers. NetLogo's run stops there.
• Recovery and resistance are decided by a real-valued draw. NetLogo draws an integer with random 100, and the two agree at whole percentages.
• average_node_degree stops at 20. NetLogo's slider reaches one less than the node count.
• virus_spread_chance and recovery_chance reach 100%, and the node count reaches 10,000,000, with 10,000 by default. NetLogo's sliders stop at 10% and 300 nodes, with 150 nodes by default.
• directed, keep_rewiring and the rewire action are additions. Rewiring follows rewire-a-link from NetLogo's Diffusion on a Directed Network, except that any unjoined pair can take the edge.

Team Assembly¶

NetLogo's Team Assembly, after Guimerà, Uzzi, Spiro and Amaral (2005). Each tick one team of team_size members is assembled, and every pair of its members is joined by an edge. A member is an incumbent with probability p and a newcomer otherwise. With probability q an incumbent member is drawn uniformly from the team's previous collaborators so far. Otherwise, or when the team has no previous collaborators, it is drawn uniformly from the incumbents outside the team. A node that goes more than max_downtime ticks without joining a team retires, and its edges go with it.

num_agents is the starting population, split into teams of team_size, each with an edge between every pair of its members. The population then grows with each newcomer and shrinks with each retirement. A larger max_downtime gives a larger steady-state population.

Newcomer-Newcomer Links, Newcomer-Incumbent Links, Incumbent-Incumbent Links and Previous Collaborator Links count the edges by type. An edge takes its type from its ends when it is made, and becomes a previous collaborator edge when the same pair meets again in a later team. Giant Component Share is the fraction of nodes in the largest connected component. Mean Component Size is the node count divided by the number of components, with an isolated node counted as its own component. These two stats are recomputed when a snapshot is published, and only if the graph has changed since the last snapshot.

Team Assembly declares no actions.

The model differs from the NetLogo version in these ways.

• NetLogo's setup is a single team. A node count equal to team_size gives the same start, as the defaults do.
• max_downtime reaches 1,000,000. NetLogo's slider stops at 100.
• When a member is to be an incumbent and no node is left outside the team, a newcomer joins instead. NetLogo stops with an error there.
• Newcomers are placed around the team's first incumbent, or around the centre of the world if the team has no incumbent. NetLogo creates them at the origin and leaves the rest to its layout.
• The layout runs in the app, outside the model, and positions are for drawing only. NetLogo's go runs its layout as part of each tick while layout? is on.
• Every node is drawn at one size. NetLogo draws the last team's members at twice the size of the other nodes.

Overriding a parameter¶

--set accepts a parameter id and a value, and can be given more than once.

cargo run --release -p henad-cli -- ants \
  --set num_agents=1000000 --set world_width=4472 --set world_height=4472 \
  --steps 1000

Keep the world area proportional to the agent count and the density holds constant, which leaves two runs at different scales comparable.