3D Mixed-Agent Ecosystem with Pathfinding
This model is much more advanced version of the Predator-prey dynamics example. It uses a 3-dimensional ContinuousSpace
, a realistic terrain for the agents, and pathfinding (with multiple pathfinders). It should be considered an advanced example for showcasing pathfinding.
Agents in this model are one of three species of animals: rabbits, foxes and hawks. Rabbits eat grass, and are hunted by foxes and hawks. While rabbits and foxes are restricted to walk on suitable portions of the map, hawks are capable of flight and can fly over a much larger region of the map.
Because agents share all their properties, to optimize performance agent types are distinguished using a type
field. Agents also have an additional energy
field, which is consumed to move around and reproduce. Eating food (grass or rabbits) replenishes energy
by a fixed amount.
using Agents, Agents.Pathfinding
using Random
import ImageMagick
using FileIO: load
@agent struct Rabbit(ContinuousAgent{3,Float64})
energy::Float64
end
@agent struct Fox(ContinuousAgent{3,Float64})
energy::Float64
end
@agent struct Hawk(ContinuousAgent{3,Float64})
energy::Float64
end
@multiagent Animal(Rabbit, Fox, Hawk)
A utility function to find the euclidean norm of a Vector
eunorm(vec) = √sum(vec .^ 2)
const v0 = (0.0, 0.0, 0.0) # we don't use the velocity field here
(0.0, 0.0, 0.0)
The environment is generated from a heightmap: a 2D matrix, where each value denotes the height of the terrain at that point. We segregate the model into 4 regions based on the height:
- Anything below
water_level
is water and cannot be walked on - The region between
water_level
andgrass_level
is flatland, that can be walked on - The part of the map between
grass_level
andmountain_level
is too high for animals to walk over, but it can be flown over - The terrain above
mountain_level
is completely inaccessible
Grass is the food source for rabbits. It can grow anywhere from water_level
to grass_level
. The spread of grass across the terrain is specified using a BitArray. A value of 1 at a location indicates the presence of grass there, which can be consumed when it is eaten by a rabbit. The probability of grass growing is proportional to how close it is to the water.
The initialize_model
function takes in the URL to our heightmap, the thresholds for the 4 regions, and some additional parameters for the model. It then creates and returns a model with the specified heightmap and containing the specified number of rabbits, foxes and hawks.
function initialize_model(
heightmap_url =
"https://raw.githubusercontent.com/JuliaDynamics/" *
"JuliaDynamics/master/videos/agents/rabbit_fox_hawk_heightmap.png",
water_level = 8,
grass_level = 20,
mountain_level = 35;
n_rabbits = 160, ## initial number of rabbits
n_foxes = 30, ## initial number of foxes
n_hawks = 30, ## initial number of hawks
Δe_grass = 25, ## energy gained from eating grass
Δe_rabbit = 30, ## energy gained from eating one rabbit
rabbit_repr = 0.06, ## probability for a rabbit to (asexually) reproduce at any step
fox_repr = 0.03, ## probability for a fox to (asexually) reproduce at any step
hawk_repr = 0.02, ## probability for a hawk to (asexually) reproduce at any step
rabbit_vision = 6, ## how far rabbits can see grass and spot predators
fox_vision = 10, ## how far foxes can see rabbits to hunt
hawk_vision = 15, ## how far hawks can see rabbits to hunt
rabbit_speed = 1.3, ## movement speed of rabbits
fox_speed = 1.1, ## movement speed of foxes
hawk_speed = 1.2, ## movement speed of hawks
regrowth_chance = 0.03, ## probability that a patch of grass regrows at any step
dt = 0.1, ## discrete timestep each iteration of the model
seed = 42, ## seed for random number generator
)
# Download and load the heightmap. The grayscale value is converted to `Float64` and
# scaled from 1 to 40
heightmap = floor.(Int, convert.(Float64, load(download(heightmap_url))) * 39) .+ 1
# The x and y dimensions of the pathfinder are that of the heightmap
dims = (size(heightmap)..., 50)
# The region of the map that is accessible to each type of animal (land-based or flying)
# is defined using `BitArrays`
land_walkmap = BitArray(falses(dims...))
air_walkmap = BitArray(falses(dims...))
for i in 1:dims[1], j in 1:dims[2]
# land animals can only walk on top of the terrain between water_level and grass_level
if water_level < heightmap[i, j] < grass_level
land_walkmap[i, j, heightmap[i, j]+1] = true
end
# air animals can fly at any height upto mountain_level
if heightmap[i, j] < mountain_level
air_walkmap[i, j, (heightmap[i, j]+1):mountain_level] .= true
end
end
# Generate the RNG for the model
rng = MersenneTwister(seed)
# Note that the dimensions of the space do not have to correspond to the dimensions
# of the pathfinder. Discretisation is handled by the pathfinding methods
space = ContinuousSpace((100., 100., 50.); periodic = false)
# Generate an array of random numbers, and threshold it by the probability of grass growing
# at that location. Although this causes grass to grow below `water_level`, it is
# effectively ignored by `land_walkmap`
grass = BitArray(
rand(rng, dims[1:2]...) .< ((grass_level .- heightmap) ./ (grass_level - water_level)),
)
properties = (
# The pathfinder for rabbits and foxes
landfinder = AStar(space; walkmap = land_walkmap),
# The pathfinder for hawks
airfinder = AStar(space; walkmap = air_walkmap, cost_metric = MaxDistance{3}()),
Δe_grass = Δe_grass,
Δe_rabbit = Δe_rabbit,
rabbit_repr = rabbit_repr,
fox_repr = fox_repr,
hawk_repr = hawk_repr,
rabbit_vision = rabbit_vision,
fox_vision = fox_vision,
hawk_vision = hawk_vision,
rabbit_speed = rabbit_speed,
fox_speed = fox_speed,
hawk_speed = hawk_speed,
heightmap = heightmap,
grass = grass,
regrowth_chance = regrowth_chance,
water_level = water_level,
grass_level = grass_level,
dt = dt,
)
model = StandardABM(Animal, space; agent_step! = animal_step!,
model_step! = model_step!, rng, properties)
# spawn each animal at a random walkable position according to its pathfinder
for _ in 1:n_rabbits
pos = random_walkable(model, model.landfinder)
agent = Animal(Rabbit(model, random_position(model), v0, rand(abmrng(model), Δe_grass:2Δe_grass)))
add_agent_own_pos!(agent, model)
end
for _ in 1:n_foxes
pos = random_walkable(model, model.landfinder)
agent = Animal(Fox(model, random_position(model), v0, rand(abmrng(model), Δe_rabbit:2Δe_rabbit)))
add_agent_own_pos!(agent, model)
end
for _ in 1:n_hawks
pos = random_walkable(model, model.airfinder)
agent = Animal(Hawk(model, random_position(model), v0, rand(abmrng(model), Δe_rabbit:2Δe_rabbit)))
add_agent_own_pos!(agent, model)
end
return model
end
initialize_model (generic function with 5 methods)
Stepping functions
The animal_step!
function dispatches to the proper function depending on the type of agent. The stepping functions for each type of agent are similar: They lose energy per step, and die if their energy ever reaches 0. They also have a random probability to reproduce at an iteration. Agents all move towards their food. In the case of rabbits, they also move away from any nearby predators.
Rabbits eat grass at their position, if it exists. If they see a predator, they run away. The direction in which they flee is dependent on all predators in their vision, with closer ones contributing more to the chosen direction. If there are no predators to flee from, rabbits walk around randomly.
animal_step!(animal, model) = animal_step!(animal, model, variant(animal))
function animal_step!(rabbit, model, ::Rabbit)
# Eat grass at this position, if any
if get_spatial_property(rabbit.pos, model.grass, model) == 1
model.grass[get_spatial_index(rabbit.pos, model.grass, model)] = 0
rabbit.energy += model.Δe_grass
end
# The energy cost at each step corresponds to the amount of time that has passed
# since the last step
rabbit.energy -= model.dt
# All animals die if their energy reaches 0
if rabbit.energy <= 0
remove_agent!(rabbit, model, model.landfinder)
return
end
# Get a list of positions of all nearby predators
predators = [
x.pos for x in nearby_agents(rabbit, model, model.rabbit_vision) if
variant(x) isa Fox || variant(x) isa Hawk
]
# If the rabbit sees a predator and isn't already moving somewhere
if !isempty(predators) && is_stationary(rabbit, model.landfinder)
# Try and get an ideal direction away from predators
direction = (0., 0., 0.)
for predator in predators
# Get the direction away from the predator
away_direction = (rabbit.pos .- predator)
# In case there is already a predator at our location, moving anywhere is
# moving away from it, so it doesn't contribute to `direction`
all(away_direction .≈ 0.) && continue
# Add this to the overall direction, scaling inversely with distance.
# As a result, closer predators contribute more to the direction to move in
direction = direction .+ away_direction ./ eunorm(away_direction) ^ 2
end
# If the only predator is right on top of the rabbit
if all(direction .≈ 0.)
# Move anywhere
chosen_position = random_walkable(rabbit.pos, model, model.landfinder, model.rabbit_vision)
else
# Normalize the resultant direction, and get the ideal position to move it
direction = direction ./ eunorm(direction)
# Move to a random position in the general direction of away from predators
position = rabbit.pos .+ direction .* (model.rabbit_vision / 2.)
chosen_position = random_walkable(position, model, model.landfinder, model.rabbit_vision / 2.)
end
plan_route!(rabbit, chosen_position, model.landfinder)
end
# Reproduce with a random probability, scaling according to the time passed each
# step
rand(abmrng(model)) <= model.rabbit_repr * model.dt && reproduce!(rabbit, model)
# If the rabbit isn't already moving somewhere, move to a random spot
if is_stationary(rabbit, model.landfinder)
plan_route!(
rabbit,
random_walkable(rabbit.pos, model, model.landfinder, model.rabbit_vision),
model.landfinder
)
end
# Move along the route planned above
move_along_route!(rabbit, model, model.landfinder, model.rabbit_speed, model.dt)
end
animal_step! (generic function with 2 methods)
Foxes hunt for rabbits, and eat rabbits within a unit radius of its position.
function animal_step!(fox, model, ::Fox)
# Look for nearby rabbits that can be eaten
food = [x for x in nearby_agents(fox, model) if variant(x) isa Rabbit]
if !isempty(food)
remove_agent!(rand(abmrng(model), food), model, model.landfinder)
fox.energy += model.Δe_rabbit
end
# The energy cost at each step corresponds to the amount of time that has passed
# since the last step
fox.energy -= model.dt
# All animals die once their energy reaches 0
if fox.energy <= 0
remove_agent!(fox, model, model.landfinder)
return
end
# Random chance to reproduce every step
rand(abmrng(model)) <= model.fox_repr * model.dt && reproduce!(fox, model)
# If the fox isn't already moving somewhere
if is_stationary(fox, model.landfinder)
# Look for any nearby rabbits
prey = [x for x in nearby_agents(fox, model, model.fox_vision) if variant(x) isa Rabbit]
if isempty(prey)
# Move anywhere if no rabbits were found
plan_route!(
fox,
random_walkable(fox.pos, model, model.landfinder, model.fox_vision),
model.landfinder,
)
else
# Move toward a random rabbit
plan_route!(fox, rand(abmrng(model), map(x -> x.pos, prey)), model.landfinder)
end
end
move_along_route!(fox, model, model.landfinder, model.fox_speed, model.dt)
end
animal_step! (generic function with 3 methods)
Hawks function similarly to foxes, except they can also fly. They dive down for prey and fly back up after eating it.
function animal_step!(hawk, model, ::Hawk)
# Look for rabbits nearby
food = [x for x in nearby_agents(hawk, model) if variant(x) isa Rabbit]
if !isempty(food)
# Eat (remove) the rabbit
remove_agent!(rand(abmrng(model), food), model, model.airfinder)
hawk.energy += model.Δe_rabbit
# Fly back up
plan_route!(hawk, hawk.pos .+ (0., 0., 7.), model.airfinder)
end
# The rest of the stepping function is similar to that of foxes, except hawks use a
# different pathfinder
hawk.energy -= model.dt
if hawk.energy <= 0
remove_agent!(hawk, model, model.airfinder)
return
end
rand(abmrng(model)) <= model.hawk_repr * model.dt && reproduce!(hawk, model)
if is_stationary(hawk, model.airfinder)
prey = [x for x in nearby_agents(hawk, model, model.hawk_vision) if variant(x) isa Rabbit]
if isempty(prey)
plan_route!(
hawk,
random_walkable(hawk.pos, model, model.airfinder, model.hawk_vision),
model.airfinder,
)
else
plan_route!(hawk, rand(abmrng(model), map(x -> x.pos, prey)), model.airfinder)
end
end
move_along_route!(hawk, model, model.airfinder, model.hawk_speed, model.dt)
end
animal_step! (generic function with 4 methods)
This function is called when an animal reproduces. The animal loses half its energy, and a copy of it is created and added to the model.
function reproduce!(animal, model)
animal.energy = Float64(ceil(Int, animal.energy / 2))
new_agent = Animal(typeof(variant(animal))(model, random_position(model), v0, animal.energy))
add_agent!(new_agent, model)
end
reproduce! (generic function with 1 method)
The model stepping function simulates the growth of grass
function model_step!(model)
# To prevent copying of data, obtain a view of the part of the grass matrix that
# doesn't have any grass, and grass can grow there
growable = view(
model.grass,
model.grass .== 0 .& model.water_level .< model.heightmap .<= model.grass_level,
)
# Grass regrows with a random probability, scaling with the amount of time passing
# each step of the model
growable .= rand(abmrng(model), length(growable)) .< model.regrowth_chance * model.dt
end
model_step! (generic function with 1 method)
Passing in a sample heightmap to the initialize_model
function we created returns the generated model.
model = initialize_model()
StandardABM with 220 agents of type Animal
agents container: Dict
space: continuous space with [100.0, 100.0, 50.0] extent and spacing=2.5
scheduler: fastest
properties: landfinder, airfinder, Δe_grass, Δe_rabbit, rabbit_repr, fox_repr, hawk_repr, rabbit_vision, fox_vision, hawk_vision, rabbit_speed, fox_speed, hawk_speed, heightmap, grass, regrowth_chance, water_level, grass_level, dt
Visualization
Now we use Makie
to create a visualization of the model running in 3D space
The agents are color-coded according to their type
, to make them easily identifiable in the visualization.
using GLMakie # CairoMakie doesn't do 3D plots well
animalcolor(a::Rabbit) = :brown
animalcolor(a::Fox) = :orange
animalcolor(a::Hawk) = :blue
animalcolor (generic function with 3 methods)
We use surface!
to plot the terrain as a mesh, and colour it using the :terrain
colormap. Since the heightmap dimensions don't correspond to the dimensions of the space, we explicitly provide ranges to specify where the heightmap should be plotted.
const ABMPlot = Agents.get_ABMPlot_type()
function Agents.static_preplot!(ax::Axis3, p::ABMPlot)
surface!(
ax,
(100/205):(100/205):100,
(100/205):(100/205):100,
p.abmobs[].model[].heightmap;
colormap = :terrain
)
end
abmvideo(
"rabbit_fox_hawk.mp4",
model;
figure = (size = (800, 700),),
frames = 300,
framerate = 15,
agent_color = animalcolor,
agent_size = 1.0,
title = "Rabbit Fox Hawk with pathfinding"
)