I'm quite new to reinforcement learning and my project will consist of detecting lanes with RL.

I'm using q-learning and I'm having a hard time thinking how my q table should look like, I mean - what could represent a state. My main idea is to feed the machine with a frame that contains a road picture, which the edge detection function is being applied to (and by thus getting lots of lines that exits in the frame). And train the machine which lines are the correct lane line. I already have a deterministic function that already recognizes the lanes and it will be the function that will teach the machine. I already organized some lane parameters such as (lane length, lane cords, lane color (white or yellow have a better probability to be a lane), lane diameter and the lane incline).

Now, my only issue is how should I construct the Q-table. Basically, what could represent a state and which lanes or decisions I should reward.

  • $\begingroup$ I'm not sure RL suits for this task, why don't you use computervision algorithms? $\endgroup$
    – malioboro
    Feb 2, 2020 at 13:52
  • $\begingroup$ RL has been used for lane keeping, which may require lane detection. Maybe have a look at the literature. $\endgroup$
    – nbro
    Feb 3, 2020 at 11:34
  • $\begingroup$ A Q-table is probably not the right tool for this problem. I would look at other RL algorithms at Spinning Up. $\endgroup$
    – S2673
    Oct 30, 2020 at 15:06

1 Answer 1


I will agree with malioboro, maybe RL is an overkill for such a task. Even though with the trend of autonomous driving research lately, papers dealing with lane changing almost certainly exist, you should check them out for more details.

As stated in Lane Change Decision-making through DeepReinforcement Learning with Rule-basedConstraints , "When the states are discrete and finite, the Q-function canbe easily formulated in a tabular form. But in many practical applications, for example, lane change decision-making task,the state space of them is very large or even continuous,using the Q-learning algorithm will lead to dimension disaster. Therefore, tabular Q-learning algorithm does not applicable to the learning problem of continuous state space and continuousaction space"

So, i would recommend using image processing techniques, or Deep-Q-Learning. Q-Learning is not capable of dealing with continuous problems like lane changing or lane tracking.


You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .