Observation Space: The Observation space for each agent is a 5×5 array that describes a list of 5 nearby vehicles by a set of features such as Position ( x, y) and Velocity (Vx, Vy) . For the purpose of Multi Agent Learning We concatenate the observation of the number of agents present into a tuple
Embedded Files
Observation space
Action Space: Our Action Space is a Discrete space which consists of Longitudanal and Lateral movement
0: 'LANE_LEFT',
1: 'IDLE',
2: 'LANE_RIGHT',
3: 'FASTER',
4: 'SLOWER'
Reward Formulation:
The vehicle is driving on a straight highway with several lanes, and is rewarded for reaching a high speed, staying on the rightmost lanes and avoiding collisions.
RIGHT_LANE_REWARD = 0.1
HIGH_SPEED_REWARD = 0.4
LANE_CHANGE_REWARD = 0
We average the reward values obtained by each agent in our learning policy. The episode ends after it reaches a time limit of 50 sec or If the agents collide.
Policy: DQN
We employ a DQN method learn Q values through a Multi Layer Perceptron Model. The actions are chosen by the policy with maximum Q value for a given state.
MLP- 2 layer network with 256*256 neurons
Page updated
Google Sites
Report abuse