invalid actions of the agent

FazelYU / Adaptive-Navigation

9 stars 0 forks source link

Closed FazelYU closed 2 years ago

FazelYU commented 3 years ago

a router may route to a road with no connection

FazelYU commented 3 years ago

subs_edge=random.choice(traci.lane.getLinks(road_ID+"0"))[0].split()[0]

FazelYU commented 3 years ago

[x] The agent must be aware of the incoming edge of AV that sends the routing query
the state dimension would be different for different agents

FazelYU commented 3 years ago

[x] restore node_dic: for each node, we should have immediate access to its outgoing and incoming edges IDs

FazelYU commented 3 years ago

FazelYU commented 3 years ago

Next state = current state / the random chosen next state/ the inavlid chosen next state Done = false

FazelYU commented 3 years ago

We can allow the U-turn for now

FazelYU commented 3 years ago

Next state = the random chosen next state

FazelYU commented 2 years ago

FazelYU commented 2 years ago

improvement: options:

[ ] 1- when the action is invalid, choose the substitute action according to the output probability distribution over actions
[x] 2- discard the invalid actions in the first place. Mask the output layer of the Q-Network over the valid actions only.

FazelYU commented 2 years ago

Done: discard the invalid actions in the first place. Mask the output layer of the Q-Network over the valid actions only.