Epsilon and learning rate decay in epsilon greedy q learning

As the answer of Vishma Dias described learning rate [decay], I would like to elaborate the epsilon-greedy method that I think the question implicitly mentioned a decayed-epsilon-greedy method for exploration and exploitation.

One way to balance between exploration and exploitation during training RL policy is by using the epsilon-greedy method. For example, =0.3 means with a probability=0.3 the output action is randomly selected from the action space, and with probability=0.7 the output action is greedily selected based on argmax(Q).

An improved of the epsilon-greedy method is called a decayed-epsilon-greedy method. In this method, for example, we train a policy with totally N epochs/episodes (which depends on the problem specific), the algorithm initially sets = (e.g., =0.6), then gradually decreases to end at = (e.g., =0.1) over training epoches/episodes. Specifically, at the initial training process, we let the model more freedom to explore with a high probability (e.g.,=0.6), and then gradually decrease the with a rate r over training epochs/episodes with the following formula:

With this more flexible choice to end at the very small exploration probability , after the training process will focus more on exploitation (i.e., greedy) while it still can explore with a very small probability when the policy is approximately converged.

You can see the advantage of the decayed-epsilon-greedy method in this post.

At the beginning, you want epsilon to be high so that you take big leaps and learn things

I think you have have mistaken epsilon and learning rate. This definition is actually related to the learning rate.

Learning rate decay

Learning rate is how big you take a leap in finding optimal policy. In the terms of simple QLearning it's how much you are updating the Q value with each step.

Higher alpha means you are updating your Q values in big steps. When the agent is learning you should decay this to stabilize your model output which eventually converges to an optimal policy.

Epsilon Decay

Epsilon is used when we are selecting specific actions base on the Q values we already have. As an example if we select pure greedy method ( epsilon = 0 ) then we are always selecting the highest q value among the all the q values for a specific state. This causes issue in exploration as we can get stuck easily at a local optima.

Therefore we introduce a randomness using epsilon. As an example if epsilon = 0.3 then we are selecting random actions with 0.3 probability regardless of the actual q value.

Find more details on epsilon-greedy policy here.

In conclusion learning rate is associated with how big you take a leap and epsilon is associated with how random you take an action. As the learning goes on both should decayed to stabilize and exploit the learned policy which converges to an optimal one.