DQN: Annealing Exploration
Beginner
The problem, from first principles
Early in training the agent knows little, so it should explore often. DQN (Mnih et al., 2015) starts with a high exploration rate and decays it over the first million frames, then holds it low.
From theory to code
Implement epsilon_schedule(step, start, end, decay_steps), a linear decay from start to end that holds end afterward.
Constraints
decay_stepsis positive.
Hints
Hint 1
Interpolate with step / decay_steps as the fraction of progress, clamped once it reaches 1.
Keywords
research-papersreinforcement-learningdeep-q-learning
Python
Run your code to see results
No results yet
Run your code to check it against the examples, or submit to run every test.