settings.ini

# This file voluntarily contains only the most important/interesting settings/hyperparameters which are defined in the .ini config files which can be found unfer the config folder.
# A description of most of these settings is provided
# You can look for the corresponding setting in a complete experiment .ini file and tune it accordingly.



[BASE]

mode = 'test' # train, test (performing training or evaluation)

# PATH USED TO LOAD PARAMETERS IF PERFORMING EVALUATION 
load_model_path : 'config/binary/GCN/Q_L/Train/DIVERSIFIED_NETS/results/BINARY_GEN_GCN_IQL/models_params'

# TYPE OF POLICY (binary is the one used in the paper)
policy : 'binary' # 'free'

save_params_frequency : 100
save_extended_training_stats:True

# DEFINES IF TRIPS ARE GENERATED BEFORE THE EXPERIMENT (GENERATING THEM BEFOREHAND ENABLES USING THE SAME TRIPS BETWEEN DIFFERENT EXPERIMENTS/METHODS)
gen_trips_before_exp : True

# Estimated arrival rate (used to influence traffic density)
period = 1

# False is faster but can cause memory overflow for large road networks and/or many self-play processes
sequential_computation : True

# Number of steps per episode
nb_steps_per_exp: 3600
# Total number of steps 
exp_sim_duration = 3600
# Maximum real-time/wallclock time duration
exp_real_duration = 1000000

# Used for debugging 
print_time_gating:False

# For evaluation, 'greedy' is used for a trained method
tests = ['greedy']   #  ['strong_baseline', 'classic']

# Number of environments running in parallel 
n_tests = 25


# When generating random networks, this value defines the proportion of road networks which will be identical to the road network used in evaluation. 
# Value between 0 and 1 
# if 0 , the method is Generalist and only trained on random road networks which do not include the target/evaluation road network)
# if 1 , the method is Specialist and is only trained on the road network used in training 
specialist : 1

# Number of times to run the entire experiment 
n_exp = 1

# Uses a real road network
real_net = True
# Adress of the real road network to be used
real_net_address : 'Manhattan.net.xml'
net_name :'Manhattan'

# Name of the experiment used to create a proper arborescence 
exp_name : 'BINARY_GEN_GCN_IQL_NO_DUELING'


[INTERFACE]

### Includes rendering (only works if code is run locally)
render : False
### Save rendering (only works if code is run locally)
save_render : False

  
[SIM]
# Speed at which vehicles enter the network
V_ENTER = 35

### WARMING
wait_n_steps : 0


yellow_duration : 5
min_time_between_actions : 5

### USED FOR DYNAMIC TRAFFIC GENERATION (NOT THE METHOD USED IN THE PAPER)
N_VEH_SAMPLES : 50
PROB_VEH : 0.003

### USED FOR GENERATING TRIPS BEFORE RUNNING SIMULATIONS (METHOD USED IN THE PAPER)
target_velocity: 50
fringe_factor: 10
demand_duration : 120
demand_variance : 0
lane_demand_variance : 2
min_distance:2
Max_Speed: 50


[MODEL]
# USE GCN OR DNN
GCN : True

[STATE]
#Lanes

# INCLUDE AGGREGATED DATA FOR LANE NODE EMBEDDINGS (FOR IG-RL-L)
lane_node_state : True 

# FEATURES TO INCLUDE IN THE LANE NODE REPRESENTATION AND THEIR CORRESPONDING DIMENSIONALITIES
lane_vars : collections.OrderedDict({'length':1, 'nb_veh' :1, 'avg_speed' : 1}) 


# FEATURES TO INCLUDE IN THE CONNECTION NODE REPRESENTATION AND THEIR CORRESPONDING DIMENSIONALITIES
connection_vars : collections.OrderedDict({'open' :1, 'nb_switch_to_open' : 1, 'current_priority' : 1, 'priority_next_open':1}) 
num_observed_next_phases : 12


# INCLUDE VEHICLE AS NODES IN THE NETWORK (FOR IG-RL-V)
veh_as_nodes : False
# FEATURES TO INCLUDE IN THE VEHICLE NODE REPRESENTATION AND THEIR CORRESPONDING DIMENSIONALITIES
veh_vars : collections.OrderedDict({traci.constants.VAR_SPEED : 1, traci.constants.VAR_LANEPOSITION : 1})


[REWARD]

# METRIC TO USE AS REWARD ('queue_length is used in the paper')
reward_type = 'queue_length' 

min_dist_delay = 25 # (length of entering area) : In training only, we add the delay of a vehicle to a lane only if it has at least traveled this value (in meters) along the corresponding lane
max_dist_queue = 50 # (length of lane sensors) : In both training and test, vehicles are counted in the queue if the distance which separates them to the intersection (in meters) is equal or inferior to this value 


# IF distance_gamma is 0, only rewards on neighbouring lanes are included in the MDP of a given intersection, if >0, other rewards will be discounted and their values will be included. (0 in the paper)
distance_gamma : 0 

# RL discount factor
time_gamma : 0.95 


# DIMENSIONALITY OF THE FINAL MLP LAYER
nn_layers_size :32

learning_rate : 1e-3
batch_size : 16
# Enables gradient accumulation between training steps (frequency at which parameters are updated)
# Set to 1 to update parameters after every gradient computation.
accumulation_steps : 1
[RL]

# FREQUENCY AT WHICH THE MODEL IN THE COMPUT PROCESS RECEIVES UPDATED PARAMETERS FROM THE TRAIN PROCESS
update_comput_model_frequency : 100

# PERFORM ACTION-CORRECTION (True for IG-RL in general. Disable for IG-RL-no_correction)
correct_actions : True 

# FREQUENCY AT WHICH THE TARGET MODEL IS UPDATED IN THE DOUBLE Q LEARNING SETTTING
target_model_update_frequency: 100
double_DQN : True
dueling_DQN : False