QCM: A QP-Based Concept Map System

1. QCM: A QP-Based Concept Map System Morteza Dehghani and Kenneth D. Forbus

2. Agenda • QR and Cognitive Science • Qualitative Concept Map System • A Modeling Tool • QP Mode • Automatic Extraction of Domain Theory and Scenario files • Bayesian Mode • Determining a Priori Probabilities using Qualitative Simulations • Examples • Related Work

3. QR and Cognitive Science • Qualitative representations capture the intuitive, causal aspects of many human mental models (Forbus & Gentner 1997) • Not handled by traditional formalisms, e.g. conditions of applicability • Qualitative modeling should become an important tool for cognitive science • Provides the formalism necessary for expressing mental models • A cognitive scientist friendly system is needed where • Qualitative representations • Explore mental models

4. Qualitative Concept Map System • Qualitative Concept Map system (QCM): • Provides a cognitive scientist friendly environment • Allows modelers: • Explore Qualitative models • Incorporate them into probabilistic models • Output/integrate them in formats usable in other forms of reasoning e.g. analogical reasoning • QCM provides a friendly interface for experimenters to enter models in two different formalisms: • Qualitative Process (QP) Theory (Forbus, 1984) • Bayesian Networks (Pearl, 1988)

5. Qualitative Concept Map System • QCM uses a concept map interface (Novak & Gowin,1984) • QCM automatically checks for any modeling errors which violate: • The laws of QP theory • The laws of probability theory • providing detailed error messages • Eg: “Increase/Decrease can only connect a process with a rate or a parameter” • Types of reasoning available in QCM: • Limit Analysis • Envisionment • Influence Resolution • Probability of evidence • Calculation of posterior-marginal

6. Qualitative Concept Map System • Graphs can be outputted in different formats: • GraphML (Brandes et al. 2001) • Predicate Calculus • Hugin (for Bayesian Networks) • QCM utilizes multiple panes to represent distinct qualitative states • important for capturing changes over time • The meta-pane allows modelers to see all the states at once • Vocabulary of specific processes and quantities can be easily extended • Expedites model creation

7. QCM: QP Mode • QP theory as a representation language for physical phenomena includes: • Continuous parameters (quantities) • Causal relationships between them (influences) • Mechanisms underlying physical causality (physical processes) • Direct influences are modeled using I+ (≡ Increases) and I- (≡ Decreases) • indicate an integral connection between two parameters • Indirect influences are modeled by ∝Q+ (≡ Influences) and ∝Q- (≡ InfluencesOpposite) • indicate functional dependence between two parameters

8. QCM: QP Mode • Our qualitative simulator (Gizmo) has been integrated into QCM. • Gizmo works as the qualitative reasoning engine of QCM • Gizmo Mk2 is a full implementation of QP theory • Gizmo has been designed to be lightweight and incremental • to be used as a module in larger systems • The user has tight control over the process of qualitative simulation in Gizmo • Algorithms for both total and attainable envisioning are included

9. Automatic Extraction of Domain Theory and Scenario files • To provide support for novice modelers, the domain theory and the scenario of the model can be automatically extracted from the graph • a major boon to novice modelers • outside of computer science, experience in writing logically quantified formulae is rare • Being able to get results without having to first write a general domain theory helps reduce the entry barrier • As their models become more complex, the automatically produced models can become a starting point

10. Automatic Extraction of Domain Theory and Scenario files • This extraction is performed by: • Going over all the nodes in the graph • for each node, determining the type of node it is (e.g. Entity, Process, Quantity). • QCM automatically obtains the required information for that type of node from the graph • The information is sent to Gizmo • If missing some required information, a detailed error message is presented to the modeler.

11. The Food Web Experiment (2007) • Using the QP mode of QCM to model psychological data • Participants were presented a scenario in nature and were asked open ended questions • Cultures involved: • Rural Menominee Native Americans, Rural European Americans • Examine the relationship between culture, expertise, and causal reasoning in the domain of biology • Qualitative Concept Maps are used for modeling and analyzing food networks • Analogical reasoning is used to perform automatic classification of the qualitative models

12. Results of the Experiment • Automatic classification of the models based on culture: • Classifying Menomonee from non-Menomonee : 64% • Automatic classification of the models based on the level of the expertise • Hunters and fishermen are considered experts within this domain • Classifying experts from non-experts within Menomonee models: 72.5%. • Classifying experts from non-experts within European American models: 52% (almost chance) • QCM was successful in capturing people’s metal models

13. Example Transcript E: Do you think that the disappearance in the bears would affect other plants and animals in the forest? S: Well, there probably be a lot more berry growth for example, because they wouldn’t be eating the berries. There probably would be a lot less maybe dead trees, because they won’t be climbing on the trees and shredding them. E: Right, and what about animals? Do you think animals would be affected? S: Yeah, like for example the animals that would be prey to the bear; the fish, for example, there’s a lot of fish that wouldn’t become prey. And I mentioned deer, specifically DEER, if deer are even bear food E: How about eagles can you think of any specific way that eagles might be affected? S: Well, eagles eat the same things that bears do, they eat fish. They might have less competition for food

14. A Sample Network Number of bears influence the eating rate there probably be a lot more berries because they wouldn’t be eating the berries Bears eat berries Number of berries increase as eating rate decreases

15. Meta-Pane

16. A Sample Network QCM of Fear (Khawf) as described in Jami 'al Sa'adat (18th century)

17. Example: Heat Transfer

18. Example: Heat Transfer Domain Theory  ;;; Processes (defprocess heat-flow :participants ((the-G :type G-type) (the-F :type F-type)) :conditions ((> (temp the-G) (temp the-F))) :quantities ((heat-flow-rate :type Rate)) :consequences ((i- (heat the-G) heat-flow-rate) (i+ (heat the-F) heat-flow-rate) (qprop (heat-flow-rate heat-flow) (temp the-G)) (qprop- (heat-flow-rate heat-flow) (temp the-F)))) ;;; Quantity Functions (defquantityfunction Rate (?thing)) (defquantityfunction heat-flow-rate (?Rate)) (defquantityfunction heat (?Amount)) (defquantityfunction Amount (?thing)) (defquantityfunction temp (?Amount)) ;;; Entities (defentity G-type :quantities ((heat :type Amount) (temp :type Amount)) :consequences ((qprop (temp G-type) (heat G-type))) :documentation "finite-thermal-physob") (defentity F-type :quantities ((heat :type Amount) (temp :type Amount)) :consequences ((qprop (temp F-type) (heat F-type))) :documentation "finite-thermal-physob")

19. Bayesian Mode • Probabilistic models are finding wide application across the cognitive and brain sciences • A method for capturing and updating uncertain beliefs about the world is needed • Bayesian networks (Pearl, 1988): provide a succinct representation for probabilities • conditional probabilities can be represented and reasoned with in an efficient manner • Providing an interface in which both QP and Bayesian formalisms can be used in parallel can potentially be helpful for cognitive scientists.

20. Bayesian Mode • Modelers can switch the mode of reasoning from QP to Bayesian and make probabilistic models. • In the Bayesian mode, modelers can perform exact inference on the network and calculate the probabilities using Recursive Conditioning (RC) (Darwiche, 2001). • Networks created in the Bayesian mode are saved in the Hugin format • The standard format for many data mining and machine learning programs

21. Determining a Priori Probabilities using Qualitative Simulations • One of the main obstacles in probabilistic reasoning: Finding the a priori probabilities of variables • One approach to overcome this obstacle is to use qualitative simulations • QCM uses the information available in the QP mode to calculate a priori probabilities of quantities used in the qualitative model • The probability distribution is defined over a set of possible worlds determined by the constraints of the qualitative model. • QCM can determine the probabilistic distribution of the values of parameters by model counting. • Can be used as a bootstrapping strategy

22. Determining a Priori Probabilities using Qualitative Simulations • The degree of belief in a statement over all the possible worlds determined by qualitative envisionment is calculated • Uniform distribution is assumed • For example: if (temp F) > (temp G) relationship is to be included as a node in the model: • QCM performs an attainable envisionment determining in how many possible worlds (temp F) β (temp G) • where β={<,<=,=,>=,>,?} • Based on this measure a probability value can be assigned to (temp F) > (temp G) • i.e. under the current constraints in n of m possible worlds (temp F) > (temp G) • the probability of (temp F) > (temp G)is n/m.

23. Example

24. Related Work • QCM is a successor to VModel (Forbus et al. 2001) • VModel was developed to help middle-school students learn science. • Vmodel uses a subset of QP theory • VModel was limited to single-state reasoning • Similar differences hold with Betty’s Brain (Biswas et al 2001) • provides a domain-specific concept map environment for students • MOBUM (Machado & Bredeweg, 2001) and VISIGARP (Bouwer & Bredeweg 2001) which have lead to Garp3 (Bredweg et al 2006, Bredweg et al 2007). • Like QCM, these environments are aimed at researchers • their focus is on constructing models for qualitative simulation • QCM focuses instead on helping capture concrete, situation specific qualitative explanations of phenomena

25. Related Work • Different approaches for qualitative Bayesian inference have been proposed: • Qualitative probabilistic networks (Wellman 1990), • Qualitative certainty networks (Parsons and Mamdani 1993) • Order of magnitude reasoning in qualitative probabilistic networks (Parsons 1995) • Keppens (2007a, 2007b) employs some of these methods for qualitative Bayesian evidential reasoning in the domain of crime investigation. • QCM integrates information available from qualitative simulations in probabilistic networks • whereas other approaches mostly use qualitative techniques in performing inference on Bayesian networks.

26. Conclusions • QCM provides the basic functionality needed for cognitive scientists to build, simulate and explore qualitative mental models • This system has been expanded in several ways since the version used in Dehghani et al. (2007): • It now uses Gizmo as its qualitative reasoning engine • Modelers can now work in a probabilistic mode • QCM automatically integrates qualitative information for calculating a priori probabilities of quantities • The interface of the system has been enhanced offering easier access to reasoning capabilities. • Models can now be exported in different formats facilitating collaboration between modelers. • QCM provides the formalism and the functionality necessary for automatic evaluation of psychological data.

27. Acknowledgments • This research was supported by a grant from the AFSOR.

28. Our Overall Approach to Culture Research MoralDM: A Cognitive model of moral decision making EA NLU Interviews, surveys,& cultural stories collected Predicate calculusrepresentations ofstories, explanations Qualitative Concept Maps

29. Our Overall Approach to Culture Research Test our hypothesis by collecting human data MoralDM: A Cognitive model of moral decision making EA NLU Interviews, surveys,& cultural stories collected Predicate calculusrepresentations ofstories, explanations Qualitative Concept Maps