The PGTG environment is highly customizable. Use these constructor arguments to modify it for your use case.
| Argument | Type | Default Value | Meaning |
|---|---|---|---|
map_path |
str or None |
None |
Path to a pregenerated map to use. If set to None a new random map is generated instead. |
random_map_width |
int |
4 |
How many tiles the random generated map is wide. |
random_map_height |
int |
4 |
How many tiles the random generated map is heigh. |
random_map_percentage_of_connections |
float |
0.5 |
How many of the possible connections between tiles are generated in the random map. Each map has a path from start to goal, so a value of 0.0 will result in minimal connections, not none. |
random_map_start_position |
tuple[int, int] or tuple[int, int, str] or "random" |
(0, -1, "west") |
Where the start of the random generated map is located. The two integers specify the x and y coordinates of the tile the start is located in, optionally its direction (“north”, “east”, “south”, or “west”) can be specified. The start can only be located on a map border. |
random_map_goal_position |
tuple[int, int] or tuple[int, int, str] or "random" |
(-1, 0, "east") |
Where the goal of the random generated map is located. The two integers specify the x and y coordinates of the tile the goal is located in, optionally its direction (“north”, “east”, “south”, or “west”) can be specified. The goal can only be located on a map border. |
random_map_minimum_distance_between_start_and_goal |
int or none |
None |
The minimal length of the path between start and goal. Can only be used if both random_map_start_position and random_map_goal_position are set to "random". The maximal allowed value is the hight plus the width of the map minus 2. |
random_map_obstacle_probability |
float in [0,1] |
0.0 |
The probability for each tile of the random map to be generated with an obstacle. |
random_map_ice_probability_weight |
float |
1.0 |
The relative probability of ice being chosen when an obstacle is generated for the random map. |
random_map_broken_road_probability_weight |
float |
1.0 |
The relative probability of broken road being chosen when an obstacle is generated for the random map. |
random_map_sand_probability_weight |
float |
1.0 |
The relative probability of sand being chosen when an obstacle is generated for the random map. |
random_map_traffic_light_probability_weight |
float |
1.0 |
The relative probability of traffic lights being chosen when an obstacle is generated for the random map. |
render_mode |
"human" or "pil_image" or None |
None |
The Gymnasium render mode. If "human" is chosen a pygame window opens and advances to the next frame whenever step() is called, "pil_image" makes render() return a PIL.Image.Image and None to no output being generated. |
features_to_include_in_observation |
list[str] |
["walls", "goals", "ice", "broken road", "sand", "traffic", "traffic_light_green", "traffic_light_yellow", "traffic_light_red"] |
What features are included in the observation as one-hot encodings. Changing this argument changes the observation space. |
use_sliding_observation_window |
bool |
False |
Wether or not the part of the map that is observable should slide along centered on the agent. If set to False the tile the agent is currently in is observed. Changing this argument changes the observation space. |
sliding_observation_window_size |
int |
4 |
How many squares the sliding observation window extends in all directions. A value of 1 results in a 3 x 3 observation windows, 2 in 5 x 5, and 3 in 7 x 7. Has no effect if use_sliding_observation_window is False. Changing this argument changes the observation space. |
use_next_subgoal_direction |
bool |
False |
Wether or not to include a additional observation that indicates the direction of the next (sub) goal in the observation. Changing this argument changes the observation space. |
sum_subgoals_reward |
int |
100 |
The total reward awarded for reaching all (sub) goals. It is equally divided among all (sub) goals. |
final_goal_bonus |
int |
0 |
Additional reward for reaching the final goal in addition to the normal (sub) goal reward defined by sum_subgoals_reward. |
crash_penalty |
int |
100 |
Penalty for moving into a wall or traffic. The value is subtracted from the reward, thus a positive value should be used. |
traffic_light_violation_penalty |
int |
50 |
Penalty for running a red light. The value is subtracted from the reward, thus a positive value should be used. |
standing_still_penalty |
int |
0 |
Penalty for not moving or accelerating each. It is applied each step. The value is subtracted from the reward, thus a positive value should be used. |
already_visited_position_penalty |
int |
0 |
Penalty for moving to a square that was already visited this episode. The value is subtracted from the reward, thus a positive value should be used. |
ice_probability |
float in [0,1] |
0.1 |
The probability of the ice obstacle triggering when moved onto and moving the agent in a random direction. |
street_damage_probability |
float in [0,1] |
0.1 |
The probability of the broken road obstacle triggering when moved onto and destroying the agents tires (thus permanently limiting its to speed to 1). |
sand_probability |
float in [0,1] |
0.2 |
The probability of the sand obstacle triggering when moved onto and halting the agent. |
traffic_density |
float in [0,1] |
0.0 |
How many percent of the squares that could have traffic are occupied by it. A value of 1.0 means that all lanes are permanently completely filled with traffic, making it impossible to avoid collisions. Values between 0.01 and 0.05 are recommended. |
traffic_light_phases_duration |
tuple[int, int, int] |
(10, 3, 10) |
Duration of the traffic light phases (green, yellow, red) in steps. The second value can be set to 0 to disable the yellow light phase. |
ignore_traffic_collisions |
bool |
False |
Wether to ignore collisions with traffic. Sometimes useful for testing. |