MaxEnt

UCL. MinNorm as an approximation of MaxEnt. • Taylor series of ln p i around p i. = k i. • Truncating at degree one and summing over i ...
3MB taille 2 téléchargements 269 vues
MinNorm approximation of MaxEnt/MinDiv problems for probability tables Patrick Bogaert and Sarah Gengler

UCL

Rebuilding probability tables

UCL

Rebuilding probability tables • Limited number of samples  Poor estimates

UCL

Rebuilding probability tables • Limited number of samples  Poor estimates

• How to integrate experts opinion ?

UCL

Rebuilding probability tables • Limited number of samples  Poor estimates

• How to integrate experts opinion ?

 Rewriting information as equality / inequality constraints

UCL

Rebuilding probability tables • Limited number of samples  Poor estimates

• How to integrate experts opinion ?

 Rewriting information as equality / inequality constraints

UCL

Rebuilding probability tables • Limited number of samples  Poor estimates

• How to integrate experts opinion ?

 Rewriting information as equality / inequality constraints • Equality constraints  MaxEnt • Inequality constraints  Minimum divergence (MinDiv)

UCL

Rebuilding probability tables • Limited number of samples  Poor estimates

• How to integrate experts opinion ?

 Rewriting information as equality / inequality constraints • Equality constraints  MaxEnt • Inequality constraints  Minimum divergence (MinDiv)  Need for an efficient methodology to rebuild probability tables from both equality and inequality constraints

UCL

The MaxEnt problem

UCL

The MaxEnt problem • Equality constraints

UCL

The MaxEnt problem • Equality constraints

• Entropy maximized subject to the equality constraints

UCL

The MaxEnt problem • Equality constraints

• Entropy maximized subject to the equality constraints

UCL

The MaxEnt problem • Equality constraints

• Entropy maximized subject to the equality constraints

 Sequence of MinNorm problems for solving the MaxEnt problem

UCL

MinNorm as an approximation of MaxEnt

UCL

MinNorm as an approximation of MaxEnt • Taylor series of ln pi around pi = ki

UCL

MinNorm as an approximation of MaxEnt • Taylor series of ln pi around pi = ki

• Truncating at degree one and summing over i

UCL

MinNorm as an approximation of MaxEnt • Taylor series of ln pi around pi = ki

• Truncating at degree one and summing over i

• In particular, if ki = 1/n

UCL

MinNorm as an approximation of MaxEnt • For any other choice of the ki ‘s, by completing the square

UCL

MinNorm as an approximation of MaxEnt • For any other choice of the ki ‘s, by completing the square

• Summing over i

Where

UCL

The MinDiv problem

UCL

The MinDiv problem • Divergence or Kullback-Leibler distance

UCL

The MinDiv problem • Divergence or Kullback-Leibler distance

• Equality constraints =0

UCL

Maximizing

The MinDiv problem • Divergence or Kullback-Leibler distance

• Equality constraints =0

Maximizing 

UCL

The MinDiv problem • Divergence or Kullback-Leibler distance

• Equality constraints =0

Maximizing  

UCL

The MinDiv problem • Divergence or Kullback-Leibler distance

• Equality constraints =0

Maximizing  

 Sequence of MinNorm problems for solving the MinDiv problem

UCL

The MinDiv problem • Divergence or Kullback-Leibler distance

• Equality constraints =0

Maximizing  

 Sequence of MinNorm problems for solving the MinDiv problem Both Equality and Inequality constraints can be processed together by MinNorm approximations

UCL

MinNorm as an approximation of MinDiv

UCL

MinNorm as an approximation of MinDiv • Taylor series around pi = ki and completing the square

UCL

MinNorm as an approximation of MinDiv • Taylor series around pi = ki and completing the square

• Summing over i

Where

UCL

Application in drainage classes mapping

UCL

Application in drainage classes mapping • Categorical data are found in a wide variety of applications

UCL

Application in drainage classes mapping

UCL

Application in drainage classes mapping • Categorical data are found in a wide variety of applications • 90 % of variables collected in soil surveys are categorical • Soil drainage, an important criterion in rating soils for various uses

UCL

Application in drainage classes mapping

UCL

Application in drainage classes mapping

UCL

Application in drainage classes mapping

UCL

Application in drainage classes mapping

UCL

Application in drainage classes mapping

UCL

Application in drainage classes mapping

UCL

Application in drainage classes mapping

UCL

Application in drainage classes mapping

UCL

Integrating the lithological map : 4 cases

UCL

Integrating the lithological map : 4 cases

UCL

Integrating the lithological map : 4 cases

UCL

Spatial prediction

UCL

Integrating the lithological map : 4 cases

UCL

Spatial prediction

UCL

Integrating the lithological map : 4 cases

UCL

Conclusions

UCL

Conclusions

UCL

Conclusions 

UCL

MinNorm Approximations

Conclusions 

UCL

MinNorm Approximations

Conclusions 

UCL

MinNorm Approximations

Conclusions 

UCL

MinNorm Approximations

Conclusions 

UCL

MinNorm Approximations

Conclusions 

UCL

MinNorm Approximations

Conclusions 

UCL

MinNorm Approximations

Thank you for your attention

UCL

References

UCL