I am working on this assignment where I made the agent learn state-action values (Q-values) with Q-learning and 100% exploration rate. The environment is the classic gridworld as shown in the following picture.

enter image description here

Here are the values of my parameters.

  • Learning rate = 0.1
  • Discount factor = 0.95
  • Default reward = 0

Reaching the trophy is the final reward, no negative reward is given for bumping into walls or for taking a step.

After 500 episodes, the arrows have converged. As shown in the figure, some states have longer arrows than others (i.e., larger Q-values). Why is this so? I don't understand how the agent learns and finds the optimal actions and states when the exploration rate is 100% (each action: N-S-E-W has 25% chance to be selected)


1 Answer 1


Q-learning is guaranteed to converge (in the tabular case) under some mild conditions, one of which is that in the limit we visit each state-action tuple infinitely many times. If your random random policy (i.e. 100% exploration) is guaranteeing this and the other conditions are met (which they probably are) then Q-learning will converge.

The reason that different state-action pairs have longer arrows, i.e. higher Q-values, is simply because the value of being in that state-action pair is higher. An example would be the arrow pointing down right above the trophy -- obviously this has the highest Q-value as the return is 1. For all other states it will be $\gamma^k$ for some $k$ -- to see this remember that a Q-value is defined as

$$Q(s, a) = \mathbb{E}_\pi \left[\sum_{j=0}^\infty \gamma^j R_{t+j+1} |S_t = s, A_t = a \right]\;;$$ so for any state-action pair that is not the block above the trophy with the down arrow $\sum_{j=0}^\infty \gamma^j R_{t+j+1}$ will be a sum of $0$'s plus $\gamma^T$ where $T$ is the time that you finally reach the trophy (assuming you give a reward of 1 for reaching the trophy).

  • $\begingroup$ Thank you for your explanation. I am also confused why certain arrows (not necessarily in the same state) should converge towards the same lengths. Why is this the case? is it because of the 100% exploration rate? $\endgroup$
    – Rim Sleimi
    Feb 20, 2021 at 21:34
  • 1
    $\begingroup$ This just means that they have the same Q-value. For your example this makes sense due to the symmetry of the board (ignoring the alien). For instance, If you take the two states on the top row and one to the left and right of the trophy, then you can reach the trophy in the same amount of steps (3, when acting optimally, and remember Q-learning converges to the optimal Q-values) from either state so they have the same Q-value. $\endgroup$
    – David
    Feb 21, 2021 at 1:10
  • $\begingroup$ I see ! thank you so much !! $\endgroup$
    – Rim Sleimi
    Feb 21, 2021 at 14:52

You must log in to answer this question.

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