EasyDeepLearn

What is the exploration-exploitation tradeoff?

easy

Answer

  • Exploitation uses the best-known action to maximize immediate reward; exploration tries new actions to discover potentially better ones.
  • Too much exploitation = premature convergence; too much exploration = wasted samples.
  • Strategies: epsilon-greedy, softmax/Boltzmann, entropy bonuses, UCB, Thompson sampling, curiosity/intrinsic rewards.
Check yourself — multiple choice
  • Exploration means always taking the greedy action
  • Epsilon-greedy explores with probability epsilon
  • Exploitation means trying random actions
  • There is no such tradeoff in modern RL

Epsilon-greedy picks a random action with prob epsilon, greedy otherwise.

#exploration

Practise Reinforcement Learning

214 interview questions in this topic.

Related questions