modelling multi-component predictive systems as petri nets

39
Modelling Multi-Component Predictive Systems as Petri Nets Manuel Martín Salvador, Marcin Budka, Bogdan Gabrys Bournemouth University, UK {msalvador, mbudka, bgabrys}@bournemouth.ac.uk ISC’2017 Warsaw, Poland May 31st, 2017

Upload: manuel-martin

Post on 29-Jan-2018

130 views

Category:

Engineering


0 download

TRANSCRIPT

Modelling Multi-Component Predictive Systems as Petri Nets

Manuel Martín Salvador, Marcin Budka, Bogdan GabrysBournemouth University, UK

{msalvador, mbudka, bgabrys}@bournemouth.ac.uk

ISC’2017Warsaw, PolandMay 31st, 2017

Outline1. Motivation2. Proposal3. Automation4. Experiments5. Conclusion and future work

MOTIVATION

Predictive modelling

LabelledData

SupervisedLearning

Algorithm

PredictiveModel

Predictive systems in the industryFault detection

Online prediction of hard-to-measure variables

Process monitoring

Demand forecasting

Classification based on computer vision

Picture is Creative Commons by Jm3

Need of preprocessingGarbage in, garbage out

Missing data

Outliers

High dimensionality

Normalisation

Lack of preprocessing can lead to inconsistent models

Multi-Component Predictive Systems

Data Postprocessing PredictionsPreprocessing Predictive Model

Multi-Component Predictive Systems

Data

Predictive Model

Postprocessing PredictionsPreprocessing Predictive Model

Predictive Model

Multi-Component Predictive Systems

Preprocessing

Data

Predictive Model

Postprocessing Predictions

Preprocessing

Preprocessing Predictive Model

Predictive Model

Requirements in the industryReliability - to provide truthful results

Robustness - to work under any circumstances or inconvenience

Transparency - to be comprehensible by human experts

Reproducibility - to replicate the results of an study

Low maintenance cost - to keep the system up-to-date at low cost

PROPOSAL

● Function composition: Difficult to model parallel paths. Can’t representate states of a system.

● Directed Acyclic Graph: Not enough to model process state or temporal behaviour..

● Petri net: Very flexible and robust mathematical background.

Expr

essi

ve p

ower

Y = h(g(f(X)))

f g hX Y

f g hX Y

How to model MCPS?

Mathematical modelling language invented in 1939 by Carl Adam Petri

token

place

transition

arc

N = (P,T,F)

Petri net

Example of Petri net

Reception Waiting Room

Check in

Consulting Room

Exit

Call in

Examination and diagnosis

Patient

Example of Petri net

Reception Waiting Room

Check in

Consulting Room

Exit

Call in

Examination and diagnosis

Example of Petri net

Reception Waiting Room

Check in

Consulting Room

Exit

Call in

Examination and diagnosis

Example of Petri net

Reception Waiting Room

Check in

Consulting Room

Exit

Call in

Examination and diagnosis

Example of Petri net

Reception Waiting Room

Check in

Consulting Room

Exit

Call in

Examination and diagnosis

Example of Petri net

Reception Waiting Room

Check in

Consulting Room

Exit

Call in

Examination and diagnosis

Example of Petri net

Reception Waiting Room

Check in

Consulting Room

Exit

Call in

Examination and diagnosis

Example of Petri net

Reception Waiting Room

Check in

Consulting Room

Exit

Call in

Examination and diagnosis

Petri nets can be more complex

Source: http://bit.ly/1XZQhYZ

A Petri net is an MCPS iff all the following conditions apply:

● The Petri net is a WRI-WF-net● The places P\{i,o} have only a single input and a single output.● The Petri net is 1-bounded.● The Petri net is 1-sound.● The Petri net is ordinary.● All the transitions with multiple inputs or outputs are AND-join or AND-split,

respectively.● Any token is a tensor (i.e. multidimensional array)

Modelling MCPS as Petri net

Example of MCPS

Classifier

o

Replace missing values

Dimensionality reduction

Outlier handling

token(data) i

place

transition

MCPS = (P, Tλ, F)

AUTOMATION

Manual● WEKA● RapidMiner● Knime● IBM SPSS

Automatic● Auto-WEKA (Bayesian optimisation)● Auto-sklearn (Bayesian optimisation + Meta-learning)● TPOT (Genetic programming)● e-Lico IDA (Ontologies + Planning)

Example of WEKA workflow

MCPS composition

What are the best algorithms to process my data?

Algorithm Selection

How to tune the hyperparameters to get the best performance?

Hyperparameter Optimisation

Combined Algorithm Selection and Hyperparameter configuration problem

k-fold cross validation

Objective function(e.g. classification error)

HyperparametersMCPSs

Training dataset

Validation dataset

Thornton, C., Hutter, F., Hoos, H.H., Leyton-Brown, K.: Auto-WEKA: combined selection and hyperparameter optimization of classification algorithms.In: Proc. of the 19th ACM SIGKDD. (2013) 847–855Martin Salvador M., Budka M., Gabrys B.: Automatic composition and optimisation of multicomponent predictive systems. IEEE Transactions on Automation Science and Engineering. under review - preprint available at https://arxiv.org/abs/1612.08789

CASH problem for MCPS

WEKA methods as search space

One-click black boxData + Time Budget → MCPS

Our contribution● Recursive extension of complex

hyperparameters in the search space.● Composition and optimisation of

MCPSs (including WEKA filters, predictors and meta-predictors)

● Petri net output as PNML format

Open-source. Download at:https://github.com/dsibournemouth/autoweka

Auto-WEKA for MCPS

WoPeD: Workflow Petri Net Designer

Open-source. Download:http://woped.dhbw-karlsruhe.de

Edit, analyze and simulate PNs

Load and save PNML

EXPERIMENTS

Building soft sensors for process industryTask: build a soft sensor to predict continuous values (regression)

7 datasets from real chemical production processes

70% training and optimisation, 30% testing

Auto-WEKA: 25 runs for 30 hours with different seeds, keep the best.

Optimisation measure: RMSE

Baseline: 4 most popular methods for soft sensors (PCR, PLS, MLP and RBF)

dataset RMSE of best (test)

Difference with baseline

absorber 0.8989 ↑ 0.0844

catalyst 0.0736 ↑ 0.1144

debutanizer 0.1745 ↓ 0.0035

drier 1.3744 ↑ 0.0573

oxeno 0.0226 ↑ 0.0042

sulfur 0.0366 ↑ 0.0030

thermalox 0.6904 ↑ 0.6170

● Data distribution can change over time and affect predictions○ External factors (e.g. weather conditions, new regulations)○ Internal factors (e.g. quality of materials, equipment deterioration)

Source: INFER project

Maintaining an MCPS

GFMMZ-Score PCA Min-Max

Tim

e

i p1 p2 p3 o

data

meta-dataprediction

[-3.1, 2.7]

x1 = 3.6

[-3.1, 3.6]

Reactive adaptation of MCPS

Conclusion and future workAutomatic composition of MCPS can speed up the process of building predictive systems though can end up being a black-box process

Representing MCPSs as Petri nets has a number of benefits:

● Increase transparency● Verification● Vendor-independent

Future work:

● Workflow algebra to model MCPSs adaptation● Timed Petri nets to model task duration and delays

THANKS!

Paper: http://bit.ly/mcps-petri-nets

Slides: http://www.slideshare.net/draxus

Manuel <[email protected]>

@draxus