Artificial intelligence in applications. Modeling, ...

Plan for the students' presentations in WiSe2024/25: 

You can select a topic from the list of available topics/materials or you can search for new materials and/or you can prepare your own example.

The dates are tentative for the presentation in the class. 

We can decide to shift them all by 1-2 weeks later. Everyone should have one presentation in the class.

- one student can work on the structure of the model and collect all information necessary to build the structure ( graph of the network)
- second student will then learn from data the probabilities for the created structure.

In the class I will show you with one automotive example (maneuver recognition for cars) how this can be done. And then you can do it in a similar way.

You can find here, in section 2. Data analysis, a number of open source data, which you can use to parameterize a Bayesian network, which you have modeled similar as you have read in the papers (see next section below) OR 

extend the existing models: maybe you have a new interpretation of the modeling, you can also combine different models in a new way or add new missing elements.

This is only a small selection of papers on road congestions.

You can search for other papers on the same topic, which discuss:

the use of Bayesian Networks for Root Cause Analysis of Traffic Congestions, Disturbances in traffic flow, traffic jam, and similar key words.

I suggest to read the paper road crash.pdf

There are also data from Australia on crash, which can be used to build a Bayesian Network and to learn  the probabilities from the data.

Diagnosis and Prediction of road congestion using Bayesian networks

Shaocheng Liu

python code for data organization

Name,  presentation date: , e-mail:

HUGIN --> Help --> Help Topics --> 

• Introduction to (Limited Memory) Influence Diagrams

• How to Build a (Limited Memory) Influence Diagram

• In part 1 of the presentation you can talk about an application of Bayesian Networks of your own choice
  • search in google/Internet about a topic or papers, e.g. you can find many applications of Bayesian networks in

Diagnostics (medical or technical) or any application in Risk Assessment, which  often use Bayesian Networks for modeling.

• See also    https://data-flair.training/blogs/bayesian-network-applications/

• Im Part 2 of your presentation you can show one example, that you have modeled and you can find more materials in

Hugin --> Help --> Help Topics (Search) --> examples

• Risk modeling with Bayesian Networks:

https://www.bayesserver.com/docs/modeling/risk/

https://library.bayesia.com/articles/#!bayesialab-knowledge-hub/2021-03-10-tech-talk-bayesia-diagnostic-app

You can install the app via the following links:

• For Android: https://play.google.com/store/apps/details?id=com.seicov.app&hl=en_US&gl=US
• For iOS: https://apps.apple.com/us/app/sei/id1534832266

chrome-extension://efaidnbmnnnibpcajpcglclefindmkaj/https://library.bayesia.com/resources/Storage/bayesialab-knowledge-hub/webinars-seminars-examples/2021-03-10-COVID-Diagnostics/TechTalk_SEICov.pdf

chrome-extension://efaidnbmnnnibpcajpcglclefindmkaj/https://library.bayesia.com/resources/Storage/bayesialab-knowledge-hub/webinars-seminars-examples/2020-03-26-COVID-19-Diagnostics/2020-03-26-Differential-Diagnosis-COVID-19.pdf

https://www.bayesia.com/bayesia/bayesialab/e-book-bayesian-networks-and-bayesialab-a-practical-introduction-for-researchers

Table of Contents 

Link to Table of Contents:

https://www.bayesia.com/bayesia/bayesialab/e-book-bayesian-networks-and-bayesialab-a-practical-introduction-for-researchers#h2_1510773337

• Overview of Book
• Chapter 1: Introduction
• Chapter 2: Bayesian Network Theory
• Chapter 3: BayesiaLab
• Chapter 4: Knowledge Modeling & Probabilistic Reasoning
• Chapter 5: Bayesian Networks and Data
• Chapter 6: Supervised Learning
  

• Chapter 7: Unsupervised Learning
• Chapter 8: Probabilistic Structural Equation Models
• Chapter 9: Missing Values Processing
• Chapter 10: Causal Effect Estimation
• Chapter 11: Causality and Optimization
• Chapter 12: Attribution, Contribution, and Counterfactuals

You can select some of the listed materials, or you can develop one model, as we  discussed on 2022-12-08 and later.

1) Topic of own interest, where you apply Bayesian Networks to model an application under uncertainties - if the application is too big, then you can split it in two parts: a) 30 min student presentation in the class; b) 20 min presentation at the exam

2) Medicine: medical diagnostics - there are many applications in medicine. 

Some very good explanations of medical applications are presented at these two webinars. There is also a PDF file to use for your presentation.

Webinar 1) Differential Diagnosis of Diseases: https://library.bayesia.com/articles/#!bayesialab-knowledge-hub/differential-diagnosis-of-diseases

Webinar 2) Differential Diagnosis of Covid-19: https://library.bayesia.com/articles/#!bayesialab-knowledge-hub/differential-diagnosis-of-covid-19

Explains how to model COVID 19 diagnostics with app for a smart phone.

Diagnostic Applications: Process Industry: Applications of object-oriented Bayesian networks for condition monitoring, root cause analysis and decision support on operation of complex continuous processes

https://www.researchgate.net/publication/223839034_Applications_of_object-oriented_Bayesian_networks_for_condition_monitoring_root_cause_analysis_and_decision_support_on_operation_of_complex_continuous_processes

If you are interested in papers, using image processing for autonomous driving,

I recommend to search on internet for publications with the following key words

"uwe franke bayesian networks and machine learning autonomous driving"

If you want, you can try to get this paper from our TH-AB library

You can select a topic from the list of available topics/materials or you can search for new materials and you can prepare your own example.

See also section 2024-05-15: Bayesian Network for Analysis and Prediction of Traffic Congestion Using the Accident Data; Student's presentation: Kranthi Talluri. You can analyse another data set and Kranthi can help you.

Reference paper

For Students in both courses 

2024-12-18: Machine Learning - Lectures & Excercises by Dr. Jamal Raiyn

Please follow the instructions on how to install WEKA on your computer, i.e.

Click the Link 'Lecture II_Part I_WEKA_Install_ 22-06-2023.pptx', in order to show/download the file.

Please, un-zip the folder - to extract all materials for this lecture and excercises.

 Short Introduction Videos (5-10 min) for self-study

WEKA - open source software tool for Data Mining, i.e. using data to make predictions

Time to learn and exploire: 5 weeks, 3-5 hours/week

email Mofeed Chaar: e01283@th-ab.de

Kranthi Talluri, 2024-12-04

2024-05-08: We discussed  slides 1 - 12 on the lecture.

2024-05-22: We continue with the lecture and do some excercises:

- Modeling of Sensor Uncertainties

- Modeling of hypotheses for manaeuver recognition in traffic

I have produced an online video from slides 30 - 66. It was published on 2022-12-08:

Continuation of the lectures and modelling of OOBN will follow on 2024-05-29. 

At the excercises we will practice with modeling of various aspects as discussed in the lectures.

In this lectures we discuss real applications from the Automobil Industry.

On 2024-05-08 we start with the applications of Object-Oriented_Bayesian_Networks_for_Detection_of_Lane_Change_Maneuvers.

This will include:

1) Problem formulation. Knowledge-based modeling represented by cognitive hypotheses. Discussion.

2) Object-orriented Bayes Nets: definition and use.

3) MODELING AND RECOGNITION OF DRIVING MANEUVERS BY THE USE OF OOBNS

   - Modeling of Sensor Data

   - Modeling of the Driving Behavior During a Lane Change Maneuver

   - Define: What is feature? Calculation of the Feature Vector

   - Define: What is hypothesis? Lane Marking Crossing Hypothesis

   - Modeling of Lane Change Maneuvers

   -  Discuss the results & Conclusions

The corresponding publications are added below as PDFs and as a Link, leading to relevant Publications on "Maneuvers Recognition".

Bei Research Gate (https://www.researchgate.net/)  you can create your own Profile and thus get access to various publications.

Object-Oriented Bayesian Networks for Detection of Lane Change Maneuvers (short conference paper)

https://www.researchgate.net/publication/249012939_Object-Oriented_Bayesian_Networks_for_Detection_of_Lane_Change_Maneuvers

Object-Oriented Bayesian Networks for Detection of Lane Change Maneuvers (long journal paper)

You can find more Details about this work here:

https://publikationen.uni-tuebingen.de/xmlui/handle/10900/49869

This topic "Bayesian Network for Analysis and Prediction of Traffic Congestion Using the Accident Data" has been also presented by a PhD Student Kranthi Talluri at the 10th international conference on Vehicle Technology and Intelligent Transport Systems (VEHITS 2024):

https://www.insticc.org/node/technicalprogram/vehits/2024/

The zip file above contains all relevant materials, data, preprocessing for modeling and the final BN model.

If you are interested, you can wokr with this data to prepare your own presentation. Alternatively you can use one of the open source data sets. Kranthi will be happy to help you with the preparatioon of data and BN model .

The original presentation was given on 

1) Learned from the data (asia.dat, and from angina.dat, attached to the Lecture - two models: angina.net and ChestClinic_gw.oobn

Try to learn it in a self study, similar as we did by the excercises. Follow the instructions in the Tutorial

http://download.hugin.com/webdocs/manuals/8.9/htmlhelp/pages/Tutorials/Structure_EM_Adaptation_Learning/EM_LearningTutorial.html

2) Use the Functionality "Joint Probability Distribution (JPD)" for the learned models, angina.net and ChestClinic_gw.oobn.

JPD is useful to identify the most probable configurations over a set of nodes when their joint is too large to be represented in main memory.

or if the joint cannot be computed.

3) Just for your information: Remark on Junction Tree and Cliques (see attachement below)

The attached models are build from expertise. Compare them with the BN models you have learned from the data, where 5% of values were missing.

Fig. 6.12 Choosing the right idiom (Neil et al. 2000)

Stau (Congestion)

Cause-consequence idiom (Was sind die Ursachen von Staus in der Stadt?)

Was wären mögliche Datenquellen, die Staus Melden? Welche Datenquellen gibt es im Internet, die Staus melden?

Daraus brauchen wir: Reconciling different views

Daraus können wir einen App für die Bürger entwerfen mit zwei Ziele:

- Informieren wo gibt es Staus in der Stadt

- Wie können Staus vermieden werden?

 The first 10 min explain how to construct BNs, based on BN fragments (idioms) as discussed in the last student's Presentation, in the paper  (Neil et al. 2000) and in chapter 6 of the eBook: Kjærulff, U. B., & Madsen, A. L. (2012). Bayesian networks and influence diagrams : A guide to construction and analysi

• The Article Building large-scale Bayesian networks - is introducing Idioms in Bayes Nets . 
• It gives you some indications  on how to build Bayesian Networks from smaller BN fragments.

     (We discuss the general principles in this lecture - as orrientation how to build your own application model.)

Semijon Mujan on 2023-05-25

- Influence Diagram

- BN Application

Look at chapter 4.2. Decision Making Under Uncertainty

Some help on how to define the Conditional Probability Tables (CPT) is provided, when  you search for:

Example 4.5 (Oil Wildcatter (Raiffa 1968)).

Example 4.7 (Oil Wildcatter Strategy).

https://moodle.th-ab.de/pluginfile.php/339820/mod_resource/content/4/ERA.pdf

This topic has been presented by Semijon Mujan on 2023-05-25

(see chapter https://moodle.th-ab.de/course/view.php?id=7231#section-32)

Naive Bayes Classifier - learning from data, without preliminary structure

Jorge Emiliano Turner Escalante 2023-12-04

Turner Escalante, Jorge Emiliano <s230359@th-ab.de>

Sebastian Nesse, Hannes Weber

2023-05-11

Note on Hypothesis (Cause) und Features (Effects):

A conclusion/decision is based on the Cause – Effect Relation.

The Hypothesis is the diagnosis (i.e., the possible cause of a problem).

How do we find that there is a problem? – Based on the symptoms of this problem, the Problem has measurable (or observable) consequences/effects.

Therefore, we need for the decision/conclusion some features (German: Merkmale), i.e., measurements, and/or observations, which appear as consequences/effects of the Problems (causes).

Thus, we ask ourselves (e.g., for the Burglary-Modell):

What did we observe/measure?

• Was an Alarm issued/sounding?
• Did Watson call due to the alarm?
• Did the Radio report a new message on earthquake?

Based on the observed/measure features we conclude/decide on the causes of a certain problem (issued alarm): Burglary or Earthquake. This means, that we meet the decision on what was the possible cause (diagnosis) of a problem, based on the evidence (observations/ measurements).

The HUGIN software tool is using the developed BN model to evaluate the hypotheses based on the concrete observations (evidences) of the features.

Therefore, for the Burglary-Modell: 

The Hypotheses are the possible causes of the problems:  Burglary or Earthquake,

and the features are the observed effects (i.e., consequences or symptoms) of the problem and its evidences: Alarm, Watson Call, Radio-Message.

Note:
       * The Hypotheses are generated from the Features. In other words: from the Features are contructed the Hypothesen.

       * The hypotheses are evaluated, based on the Evidence (=measured Features) and this allows to make the decision.

       The Hypotheses are used in two different types of decision tasks:

1. Diagnosis:

           For Diagnostics: die decision finds what Hypothesis has the highest probability, based on the observations/measurements (evidence). An example of Diagnostics task is the angina.oobn Bayes Network.

      2. Prognosis:  

To download the BN examples (click on the icon).

I have collected here some examples of BN to help your preparation for the exam. You can select one of these examples or you can also present your own example (or another known example).

During the exam - you can explain the relevant exam questions based on your selected BN. 

Sucessfull Preparations for the exam! 

• //  (In German: Einige Bayes Netzwerke zu verschiedene Entscheidungsfindung Beispiele.

  Sie können ein Bayes Netz aus den eingefügten Beispiele wählen oder ein eigenen (oder ein bekannten)  Beispiel verwenden.

• Bei der Prüfung können Sie die Prüfungsfragen anhand eines Beispiels erläutern.

•  Viel Erfolg bei der Vorbereitung!

We can discuss all possible question about the exam at our last lecture+excercise  on 12 Januar.

If needed, we can also arrange an aditional meeting. We talk on January 12th. about suitable times.

Bemerkung über Hypothese (Ursache/Cause) und Merkmale (Wirkung/Effect):

Eine Entscheidungsfindung basiert auf der Ursache – Wirkung Beziehung (= Cause – Effect Relation)

Die Hypothese ist die Diagnose (d.h. die mögliche Ursachen eines Problems).

Also, wir fragen uns (z.B. bei dem Burglary-Modell):

Was hat man beobachtet?

• Hat sich ein Alarm ausgelöst?
• Hat Watson angerufen deswegen?
• Hat das Radio neue Nachrichten zur Lage (=Erdbeben) berichtet?

Anhand der Merkmale schließen wir zurück auf die Ursache eines Problems (Alarm Auslösung): Einbruch oder Erdbeben, d.h. wir treffen die Entscheidung über was war die mögliche Ursache (=Diagnose). Dabei werden die Hypothesen ausgewertet anhand der konkrete Beobachtungen (Evidenzen) der Merkmalen.

---