mcmc in structure space mcmc in order space

39

Upload: theodore-eaton

Post on 18-Jan-2016

234 views

Category:

Documents


1 download

TRANSCRIPT

Page 1: MCMC in structure space MCMC in order space
Page 2: MCMC in structure space MCMC in order space
Page 3: MCMC in structure space MCMC in order space
Page 4: MCMC in structure space MCMC in order space
Page 5: MCMC in structure space MCMC in order space
Page 6: MCMC in structure space MCMC in order space
Page 7: MCMC in structure space MCMC in order space

MCMC in structure space

Page 8: MCMC in structure space MCMC in order space

MCMC in order space

Page 9: MCMC in structure space MCMC in order space
Page 10: MCMC in structure space MCMC in order space
Page 11: MCMC in structure space MCMC in order space
Page 12: MCMC in structure space MCMC in order space
Page 13: MCMC in structure space MCMC in order space

Current work with Marco Grzegorczyk

• MCMC in structure rather than order space.

• Design new proposal moves that achieve faster mixing and convergence.

Page 14: MCMC in structure space MCMC in order space

First idea

Propose new parents from the distribution:

•Identify those new parents that are involved in the formation of directed cycles.

•Orphan them, and sample new parents for them subject to the acyclicity constraint.

Page 15: MCMC in structure space MCMC in order space
Page 16: MCMC in structure space MCMC in order space

Problem: This move is not reversible

Page 17: MCMC in structure space MCMC in order space

Design a complementary backward move, which proposes “illegal” strucures.

•Select a node X.

•Select a subset of its parents, propose new parents for these parents such that you get directed cycles that involve node X.

•Orphan node X, then select new parents subject to the acylicity constraint.

Page 18: MCMC in structure space MCMC in order space
Page 19: MCMC in structure space MCMC in order space

Move reversible, but maths complicated

Page 20: MCMC in structure space MCMC in order space

Devise a simpler move with similar mixing and convergence

•Identify a pair of nodes X Y

•Orphan both nodes.

•Sample new parents from the Boltzmann distribution subject to the acyclicity constraint such the inverse edge Y X is included.

Page 21: MCMC in structure space MCMC in order space
Page 22: MCMC in structure space MCMC in order space

This move is reversible!

Page 23: MCMC in structure space MCMC in order space
Page 24: MCMC in structure space MCMC in order space

Acceptance probability

Page 25: MCMC in structure space MCMC in order space

• Does the new method avoid the bias intrinsic to order MCMC?

• How do convergence and mixing compare to structure and order MCMC?

• What is the effect on the network reconstruction accuracy?

Page 26: MCMC in structure space MCMC in order space

• Does the new method avoid the bias intrinsic to order MCMC?

• How do convergence and mixing compare to structure and order MCMC?

• What is the effect on the network reconstruction accuracy?

Page 27: MCMC in structure space MCMC in order space

Estimating the bias of the method

• Consider a small network with only five nodes.

• A complete enumeration of structures is possible to compute the correct posterior distribution.

• Compute the difference between the predicted and the true marginal posterior probability, for all edges

Page 28: MCMC in structure space MCMC in order space
Page 29: MCMC in structure space MCMC in order space
Page 30: MCMC in structure space MCMC in order space

• Does the new method avoid the bias intrinsic to order MCMC?

• How do convergence and mixing compare to structure and order MCMC?

• What is the effect on the network reconstruction accuracy?

Page 31: MCMC in structure space MCMC in order space

Alarm network

Devised by Beinlich et al., 1989N=37 nodes

46 directed edges

We generated data sets with m=25,50,100,250,500,750,1000

instances

Page 32: MCMC in structure space MCMC in order space
Page 33: MCMC in structure space MCMC in order space
Page 34: MCMC in structure space MCMC in order space
Page 35: MCMC in structure space MCMC in order space

• Does the new method avoid the bias intrinsic to order MCMC?

• How do convergence and mixing compare to structure and order MCMC?

• What is the effect on the network reconstruction accuracy?

Page 36: MCMC in structure space MCMC in order space

• Does the new method avoid the bias intrinsic to order MCMC?

• How do convergence and mixing compare to structure and order MCMC?

• What is the effect on the network reconstruction accuracy?

Page 37: MCMC in structure space MCMC in order space
Page 38: MCMC in structure space MCMC in order space

• The new method avoids the bias intrinsic to order MCMC.

• Its convergence and mixing are similar to order MCMC; both methods outperform structure MCMC.

• Its network reconstruction accuracy is similar to order MCMC; both methods outperform structure MCMC.

• We expect to get an improvement over order MCMC when using explicit prior knowledge.

Conclusions

Page 39: MCMC in structure space MCMC in order space

Thank you