Forest Fire Environment Maker for OpenAI Gym.
gym_forest_fire
version 2.4
- Install OpenAi Gym
pip install gym
- Download and install
gym-forest-fire
git clone https://github.com/elbecerrasoto/gym-forest-fire
cd gym-forest-fire
pip install -e .
- Import and have fun
import gym_forest_fire
from gym_forest_fire import EnvMakerForestFire, ForestFire
The package implements a tool for making environments
for Reinforcement Learning tasks.
It also contains a preloaded environment and a forest fire cellular automaton simulator.
The created environments follow the guidelines of the Open AI gym API, and would contain the following methods:
from gym_forest_fire import EnvMakerForestFire
env = EnvMakerForestFire()
env.reset()
action = 7
env.step(action)
env.render()
env.close()
It also contains two extra methods for generating actions, from a random or heuristic policy. The heuristic merely follows the fire, when the CA is not updating it drives the helicopter to any fire cell, if it is updating it goes to a tree cell with fire neighbors, random otherwise.
from gym_forest_fire import EnvMakerForestFire
env = EnvMakerForestFire()
obs = env.reset()
action_random = env.random_policy()
action_heuristic = env.heuristic_policy()
obs, reward, done, info = env.step(action_heuristic)
The created environment simulates a helicopter trying to extinguish a forest fire.
The forest is simulated using a Forest Fire Automaton [Drossel and Schwabl (1992)] and
the helicopter as a position on top of the lattice and some effect over the cells.
At each time step the Agent has to make a decision to where in the neighborhood to move the helicopter,
then the helicopter moves and has some influence over the destination cell,
the effect is simply changing it to another cell type, usually from 'fire' to 'empty'
and the reward is some function of the current state of the system,
usually just counting cells types, multiplying for some weights (positive for trees and negative for fires) and adding up.
Forest Fire Automaton Drossel and Schwabl (1992)
Three type of cells: TREE, EMPTY and FIRE.
At each time step and for each cell apply the following rules
(order does not matter).
- Lighting Rule
With probability f:
TREE turns into Fire - Propagation Rule
If at least one neighbor is FIRE:
TREE turns into Fire - Burning Rule
Unconditional:
FIRE turns into EMPTY - Growth Rule
With probability p:
EMPTY turns into TREE
Also two more cells were added.
- ROCK, does not interacts with anything Used as a true death cell Used on the Deterministic mode Used on the invariant boundary conditions
- LAKE, does not interacts with anything Used on other classes that inherit from ForestFire
Deterministic mode: The automaton does not computes the Lighting and Growth rules, stops when there are no more FIRE cells.
For an overview of all the capabilities of EnvMakerForestFire check the documentation.
help(EnvMakerForestFire)
The package also contains an already instantiated example of an environment, described below.
The ForestFire-v0 environment implements a forest fire cellular automaton of 16x16 cells, with parameters f=0.005 and p=0.100
The control task is to move a helicopter through the lattice, to try to extinguish the fire. The helicopter has the effect of turning fire cells to empty cells whenever on top of them.
The possible actions to take are 9, either moving one-step into 8 directions, or staying at the same place.
Each number from 1 to 9 represents one direction.
- Left-Up
- Up
- Right-Up
- Right
- Don't move
- Left
- Left-Down
- Down
- Right-Down
The helicopter can move 8 times before the next computation of the forest fire automaton, then moving 1 more time after the updating. This is iterated over and over until termination. Basically, the helicopter can travel half the distance of the forest before the next actualization. This roughly represents the helicopter's speed.
The reward scheme is -1 per burning tree at each time.
The task is continuing.
The representation of the lattice is a numpy matrix.
- 0.77 represents a tree
- 0.66 represents empty
- -1.0 represents fire
The observation returned by the the step method is a tuple of three elements, the first is the lattice the second element is the postion of the helicopter in a [row, col] format, and the third the remaining moves of the helicopter to update the cellular automaton.
The starting position of the helicopter is 8,8, just in the middle. The starting forest configuration is random, with 0.75 chance of a tree and 0.15 of an empty space.
The cell numeration starts from the left and upper corner. So the cell at 0,0 is at the most left and upper postion and the cell at 15,15 is at most right and down postion.
Start by importing the package and initializing the environment
import gym
import gym_forest_fire
env = gym.make('ForestFire-v0')
Implementing the random policy
import numpy as np
import gym
help(EnvMakerForestFire)
env = gym.make('ForestFire-v0')
# First observation
obs = env.reset()
env.render()
total_reward = 0
for i in range(128):
print('.', end='')
action = env.random_policy()
obs, reward, done, info = env.step(action)
total_reward += reward
env.render()
print('\nTotal Reward: {}'.format(total_reward))
import gym_forest_fire
import numpy as np
symbols = {'tree':0.77, 'empty':0.66, 'fire':-1.0, 'rock':0.88, 'lake':0.99}
rows = 7
cols = 7
r_mid = int(rows/2)
c_mid = int(cols/2)
# All rocks grid
grid = np.repeat(symbols['rock'],rows*cols).reshape(rows, cols)
grid[r_mid,:] = symbols['tree']
grid[:, c_mid] = symbols['tree']
grid[0,:] = symbols['tree']
grid[r_mid, c_mid] = symbols['rock']
grid[2, c_mid-1] = symbols['tree']
grid[2, c_mid] = symbols['rock']
env = EnvMakerForestFire(init_pos_row=r_mid, init_pos_col=c_mid, custom_grid=grid,
p_fire=0.01, p_tree=0.75,
sub_tree='empty',
moves_before_updating=2,
reward_tree=3.0, reward_fire=-1.0,
**symbols)
env.reset()
env.render()