- author = "J. Rosca and D. H. Ballard",
- title = "Evolution-based discovery of hierarchical behaviors",
- booktitle = "Proceedings of the Thirteenth National Conference on Artificial Intelligence (AAAI-96)",
- year = "1996",
- publisher = "AAAI / The MIT Press",
- keywords = "genetic algorithms, genetic programming",
-
URL = "
ftp://ftp.cs.rochester.edu/pub/u/rosca/gp/96.aaai.ps.gz",
-
URL = "
https://www.aaai.org/Papers/AAAI/1996/aaai96-132.php",
- size = "7 pages",
- abstract = "The complexity of policy learning in a reinforcement learning task deteriorates primarily with the increase of the number of observations. Unfortunately, the number of observations may be unacceptably high even for simple problems. In order to cope with the scale up problem we adopt procedural representations of policies. Procedural representations have two advantages. First they are implicit, allowing for good inductive generalization over a very large set of input states. Second they facilitate modularization. In this paper we compare several randomized algorithms for learning modular procedural representations. The main algorithm, called Adaptive Representation through Learning (ARL) is a genetic programming extension that relies on the discovery of subroutines. ARL is suitable for learning hierarchies of subroutines and for constructing policies to complex tasks. When the learning problem cannot be solved because the specification is too loose and the domain is not well understood, ARL will discover regularities in the problem environment in the form of subroutines, which often lead to an easier problem solving. ARL was successfully tested on a typical reinforcement learning problem of controlling an agent in a dynamic and non-deterministic environment where the discovered subroutines correspond to agent behaviors.",
-
notes = "
Pac-Man ARL {"}differential fitness and block activation heuristics{"} {"}two distinct tiers{"} typed, random search, simulated annealing (PSA), hand-coding, See also \cite{rosca:1996:video}",
