Noisy nets — how do they replace ε-greedy?
hardAnswer
- Add parametric noise to weights of last layer(s): W = μ + σ ⊙ ε with learnable μ, σ.
- Exploration comes from state-dependent noise, learned per-parameter.
- As σ shrinks, exploration decreases automatically → no manual ε schedule.
- Used in Rainbow DQN.
- Advantage: more directed exploration than uniform ε-greedy; noise magnitude adapts to what the network needs.
Check yourself — multiple choice
- Same as epsilon-greedy
- Parametric noise W = μ + σ ⊙ ε on weights, learnable μ + σ → state-dependent exploration; σ shrinks as learning progresses; replaces manual ε schedule
- Random
- Not real
Noisy nets: parametric exploration via noisy weights; replaces ε-greedy.
#exploration#deep-rl
Practise Reinforcement Learning
214 interview questions in this topic.