Strength or Accuracy
- Exclusive ContentScan the cover in the app
- CommunityDiscuss & Share
About this book
Learning classifier systems offer a flexible way to tackle many machine learning problems. They build condition-action rules automatically and then judge how useful each rule is. In reinforcement learning, they tackle two hard problems at once: finding a good policy and generalising that policy, along with related structures such as value functions. For more than two decades the approach produced uneven results, and the causes were often obscure. In 1995, Stewart Wilson introduced XCS, which he presented as the long-awaited breakthrough. Its most important departure from earlier designs is how it values rules for the genetic algorithm that generates them. Like most classifier systems, XCS uses a genetic algorithm to create rules, but it scores rule fitness differently. Wilson called XCS accuracy-based and the earlier systems strength-based. In strength-based systems, a rule's fitness tracks the reward it collects. In XCS, fitness depends on how accurately the rule predicts that reward. The distinction is really about credit assignment, meaning how much of the system's performance each rule is considered responsible for. XCS is a Q-learning system and extends tabular Q-learning, since its rules group together states and actions. Like other Q-learners, it uses Q-values to guide which action it selects.
Reader reviews
No reviews yet. Share your thoughts on this book in the Appilog app.
Book Details
- ISBN
- 9781852337704
- Author
- Tim Kovacs
- Publisher
- Unknown Publisher
- Genres
- Technology & Computing
Media & Extras
This book has 0 videos that play on its cover. Scan the cover with the Appilog app to watch.
Get the app

