Keras-RL provides an -greedy Q Policy called rl.policy.EpsGreedyQPolicy that we can use to balance exploration and exploitation. We can use rl.policy.LinearAnnealedPolicy to decay our  as the agent steps forward in the world, as shown in the following code:

policy = LinearAnnealedPolicy(EpsGreedyQPolicy(), attr='eps', value_max=1., value_min=.1, value_test=.05, nb_steps=10000)

Here we're saying that we want to start with a value of 1 for  and go no smaller than 0.1, while testing if our random number is less than 0.05. We set the number of steps between 1 and .1 to 10,000 and Keras-RL handles the decay math for us.