next up previous
Next: Delayed Reward Up: Immediate Reward Previous: REINFORCE Algorithms

Logic-Based Methods

Another strategy for generalization in reinforcement learning is to reduce the learning problem to an associative problem of learning boolean functions. A boolean function has a vector of boolean inputs and a single boolean output. Taking inspiration from mainstream machine learning work, Kaelbling developed two algorithms for learning boolean functions from reinforcement: one uses the bias of k-DNF to drive the generalization process [54]; the other searches the space of syntactic descriptions of functions using a simple generate-and-test method [53].

The restriction to a single boolean output makes these techniques difficult to apply. In very benign learning situations, it is possible to extend this approach to use a collection of learners to independently learn the individual bits that make up a complex output. In general, however, that approach suffers from the problem of very unreliable reinforcement: if a single learner generates an inappropriate output bit, all of the learners receive a low reinforcement value. The CASCADE method [52] allows a collection of learners to be trained collectively to generate appropriate joint outputs; it is considerably more reliable, but can require additional computational effort.

Leslie Pack Kaelbling
Wed May 1 13:19:13 EDT 1996