#### DMCA

## Probabilistic Reasoning Through Genetic Algorithms and Reinforcement Learning

### Citations

8886 |
Probabilistic Reasoning in Intelligent Systems: Networks of Plausible Inference
- Pearl
- 1988
(Show Context)
Citation Context ...ng controller to direct a genetic algorithm. The random variables of a Bayesian network can be grouped into several sets reflecting the strong probabilistic correlations between random variables in the group. We build a reinforcement learning controller to identify these groups and recommend the use of "group" crossover and "group" mutation for the genetic algorithm based on these groupings. The system then evaluates the performance of the genetic algorithm and continues with reinforcement learning to further tune the controller to search for a better grouping. Introduction Bayesian Networks (Pearl 1988a) are one of the most popular models for uncertainty. Knowledge is organized in a hierarchical fashion providing easy visualization of the reasoning domain. Such networks consist of directed acyclic graphs of nodes, each representing a random variable(rv) with a finite domain. The directed arcs between the nodes represent probabilistic conditional dependencies. The joint probability over the random variables can be computed via the chain rule and the given conditional independence assumption. Bayesian networks have been applied to various domains such as story comprehension, planning, circuit... |

140 |
Finding MAPs for belief networks is NP-hard
- Shimony
- 1994
(Show Context)
Citation Context ...e joint probability over the random variables can be computed via the chain rule and the given conditional independence assumption. Bayesian networks have been applied to various domains such as story comprehension, planning, circuit fault detection and medical diagnoses. There are two types of computations performed with Bayesian networks: belief updating and belief revision(Pearl 1988b). Belief updating concerns the computation of probabilities over random variables, while belief revision concerns finding the maximally probable global assignment. However, both tasks are known to be NP-hard (Shimony 1994). In this paper, we demonstrate how a kind of genetic algorithms directed by a reinforcement learning controller can be effectively used in belief revision. A Genetic Algorithm(GA) is an evolutionary computation technique inspired from the principles of natural selection to search a solution space. Most researches modified their implementation of GA either by using non-standard chromosome representation or by designing problem specific genetic operations(Michalewizs 1996) to accommodate the problem to be solved, thus building efficient evolution programs. In this paper, we employ "group" cross... |

66 | Procedural help in Andes: Generating hints using a Bayesian network student model. - Gertner, Conati, et al. - 1998 |

26 |
GALGO: A Genetic ALGOrithm decision support tool for complex uncertain systems modeled with bayesian belief networks.
- Rojas-Guzman, Kramer
- 1993
(Show Context)
Citation Context ... on interpretation of e. The problem is to find an explanation w* such that: p(w*) = maxwew p(wle) Intuitively, we can think of the non-evidence rvs in W as possible hypotheses for e. With a small network, a valid solution method is to simply tabulate all the possible values of the rvs and then calculate the probabilities. Once the network gets larger and more complex, this method is obviously unacceptable and more efficient method must be employed. Several people have used GAs to perform belief revision over Bayesian network, but only superficially considered the topological structure of BN (Rojas-Guzman & Kramer 1993; Santos, Shimony, & Williams 1997; Santos & Shimony 1998; Welch 1996). From experimental results, we know that even the topological structure will effect the performance of GA (Williams, Santos, Shimony 1997; Jitnall & A.E.Nicholson 1996). So we use reinforcement learning to continuously learn and identify the problem-specific attributes of the BN, and employ this in the GA through "group" crossover and "group" mutations for the search process. Copyright © 1999, American Association f rArtificial Intelligence (www.aaai.org). All ights reserved. UNCERTAIN REASONING 477 From: Proceedings of the... |

7 |
Real time estimation of bayesian networks
- Welch
- 1996
(Show Context)
Citation Context ...maxwew p(wle) Intuitively, we can think of the non-evidence rvs in W as possible hypotheses for e. With a small network, a valid solution method is to simply tabulate all the possible values of the rvs and then calculate the probabilities. Once the network gets larger and more complex, this method is obviously unacceptable and more efficient method must be employed. Several people have used GAs to perform belief revision over Bayesian network, but only superficially considered the topological structure of BN (Rojas-Guzman & Kramer 1993; Santos, Shimony, & Williams 1997; Santos & Shimony 1998; Welch 1996). From experimental results, we know that even the topological structure will effect the performance of GA (Williams, Santos, Shimony 1997; Jitnall & A.E.Nicholson 1996). So we use reinforcement learning to continuously learn and identify the problem-specific attributes of the BN, and employ this in the GA through "group" crossover and "group" mutations for the search process. Copyright © 1999, American Association f rArtificial Intelligence (www.aaai.org). All ights reserved. UNCERTAIN REASONING 477 From: Proceedings of the Twelfth International FLAIRS Conference. Copyright © 1999, AAAI (www.... |

6 | A reinforcement learning neural networks for adaptive control of markhov chains. - Santharazn, Sastry - 1997 |

6 | Deterministic approximation of marginal probabilities in bayes nets.
- Santos, Shimony
- 1998
(Show Context)
Citation Context ... w* such that: p(w*) = maxwew p(wle) Intuitively, we can think of the non-evidence rvs in W as possible hypotheses for e. With a small network, a valid solution method is to simply tabulate all the possible values of the rvs and then calculate the probabilities. Once the network gets larger and more complex, this method is obviously unacceptable and more efficient method must be employed. Several people have used GAs to perform belief revision over Bayesian network, but only superficially considered the topological structure of BN (Rojas-Guzman & Kramer 1993; Santos, Shimony, & Williams 1997; Santos & Shimony 1998; Welch 1996). From experimental results, we know that even the topological structure will effect the performance of GA (Williams, Santos, Shimony 1997; Jitnall & A.E.Nicholson 1996). So we use reinforcement learning to continuously learn and identify the problem-specific attributes of the BN, and employ this in the GA through "group" crossover and "group" mutations for the search process. Copyright © 1999, American Association f rArtificial Intelligence (www.aaai.org). All ights reserved. UNCERTAIN REASONING 477 From: Proceedings of the Twelfth International FLAIRS Conference. Copyright © 199... |

5 | Experiments with distributed anytime inferencing: Working with cooperative algorithms. - Williams, Santos, et al. - 1997 |

2 | A user's guide to GENESIS Version 5.0. Navy Center for Applied Research - Grefenstette - 1990 |

2 |
Genetic Algorithms + Data Structure = Evolution Programs
- Michalewizs
- 1996
(Show Context)
Citation Context ..., while belief revision concerns finding the maximally probable global assignment. However, both tasks are known to be NP-hard (Shimony 1994). In this paper, we demonstrate how a kind of genetic algorithms directed by a reinforcement learning controller can be effectively used in belief revision. A Genetic Algorithm(GA) is an evolutionary computation technique inspired from the principles of natural selection to search a solution space. Most researches modified their implementation of GA either by using non-standard chromosome representation or by designing problem specific genetic operations(Michalewizs 1996) to accommodate the problem to be solved, thus building efficient evolution programs. In this paper, we employ "group" crossover and "group" mutation based on grouping random variables of a Bayesian Network(BN). Experimental results show that different groupings effect the performance of the GA. We use a Reinforcement Learning(RL) controller to adaptively determine the elements in each group. This method investigates the impact of using domain knowledge during the recombination and mutation phases of the GA. Bayesian Networks and Belief Revision Belief revision is the process of determining th... |

1 | Belief xmtwork inference algorithms: a study of performance based on domain characterisation. - Jitnall, Nicholson - 1996 |

1 | Hybrid algorithms for approximate belief updating in bayes nets. - Smltos, Shinmny, et al. - 1997 |