Learning Naive Physics Models and Misconceptions Scott E. Friedman (friedman@northwestern.edu) Qualitative Reasoning Group, Northwestern University, 2133 Sheridan Rd Evanston, IL 60208 USA Kenneth D. Forbus (forbus@northwestern.edu) Qualitative Reasoning Group, Northwestern University, 2133 Sheridan Rd Evanston, IL 60208 USA We next briefly summarize the relevant aspects of qualitative process theory and structure-mapping theory used in the simulation. Then we describe how our stimuli are represented and encoded, motivated by results and ideas from the cognitive science literature (diSessa, 1993; Talmy, 1988; Zacks, Tversky, & Iyer, 2001). The learning process itself is described next, followed by how the learned models are used in reasoning. We show that the simulation’s explanations of a situation where a book is at rest on a table are compatible with student explanations (Brown, 1994). We close with other related work and future work. Abstract Modeling how intuitive physics concepts are learned from experience is an important challenge for cognitive science. We describe a simulation that can learn intuitive causal models from a corpus of multimodal stimuli, consisting of sketches and text. The simulation uses analogical generalization and statistical tests over qualitative representations it constructs from the stimuli to learn abstract models. We show that the explanations the simulation provides for a new situation are consistent with explanations given by naive students. Keywords: Cognitive modeling; conceptual change; misconceptions; naive physics; qualitative reasoning Background Introduction Many people have intuitive models of physical domains that are at odds with scientific models (Minstrell, 1982; McCloskey, 1983; diSessa, 1993; Brown, 1994). While productive for reasoning about everyday physical phenomena, these naive models cause patterns of misconceptions. These misconceptions may result from improperly generalizing or contextualizing experience (Smith, diSessa, & Roschelle, 1994) or from incorporating instruction into a flawed intuitive framework (Vosniadou, 1994). Understanding how such intuitive models come about is an important problem for understanding how people learn physical domains (Forbus & Gentner, 1986). We believe it is important for computational models of domain learning and conceptual change (e.g. Ram, 1993; Esposito et al., 2000) to encompass the learning of the initial intuitive concepts. This paper describes a simulation of learning intuitive physics models from experience. Experiences are provided as combinations of sketches and natural language, which are automatically processed to produce symbolic representations for learning. The system identifies and encodes instances of the concepts to be learned and constructs qualitative representations of behavior across time. Analogical generalization is used with a statistical criterion to induce abstract models of typical patterns of behavior, which constitutes our representation of intuitive models. These models can be used to make predictions and perform simple counterfactual reasoning. We compare its explanations to those of human students on a simple reasoning task (Brown 1994). People’s intuitive physical knowledge appears to rely heavily on qualitative representations (Forbus & Gentner, 1986; Baillargeon, 1998). Consequently, we use qualitative process theory (Forbus, 1984) as part of our model. In QP theory, physical processes are the mechanism of causality for changes in dynamic systems. However, the learning we are describing here is what provides the foundation for ultimately learning physical processes; in the framework of (Forbus & Gentner, 1986), we are modeling the construction of protohistories from experience, and building on those a causal corpus consisting of causal relationships between those typical patterns of behavior. To model these patterns of behavior, we use the concept of encapsulated history (EH) from QP theory. An encapsulated history represents a category of abstracted behavior, over some span of time. It can include multiple qualitative states and events. Encapsulated histories are used when a learner does not yet understand how to reduce a behavior to physical processes. Encapsulated histories are a type of schema, and consequently have variables. The participants are the entities that an EH is instantiated over. The conditions are statements which must be true for an instance of the EH to be active. When an instance of an EH is active, the statements in its consequences are assumed to be true. Encapsulated histories are a form of explanatory schema: When instantiated, they provide an explanation for a behavior via recognizing it as an instance of a typical pattern, and furthermore can provide causal explanations, if there is causal information in the consequences.
[1]
M. McCloskey.
Naive Theories of Motion.
,
1982
.
[2]
J. Minstrell.
Explaining the ’’at rest’’ condition of an object
,
1982
.
[3]
Dedre Gentner,et al.
Structure-Mapping: A Theoretical Framework for Analogy
,
1983,
Cogn. Sci..
[4]
P. Johnson-Laird.
Mental models
,
1989
.
[5]
Kenneth D. Forbus.
Qualitative Process Theory
,
1984,
Artif. Intell..
[6]
John R. Anderson,et al.
MACHINE LEARNING An Artificial Intelligence Approach
,
2009
.
[7]
Kenneth D. Forbus,et al.
Learning Physical Domains: Toward a Theoretical Framework.
,
1986
.
[8]
Leonard Talmy,et al.
Force Dynamics in Language and Cognition
,
1987,
Cogn. Sci..
[9]
David E. Brown,et al.
Overcoming misconceptions via analogical reasoning: abstract transfer versus explanatory model construction
,
1989
.
[10]
Brian Falkenhainer,et al.
The Structure-Mapping Engine: Algorithm and Examples
,
1989,
Artif. Intell..
[11]
A. Ram.
Creative Conceptual Change
,
1993
.
[12]
A. diSessa.
Toward an Epistemology of Physics
,
1993
.
[13]
John J. Clement,et al.
Preconceptions in mechanics : lessons dealing with student' conceptual difficuties
,
1994
.
[14]
J. Roschelle,et al.
Misconceptions Reconceived: A Constructivist Analysis of Knowledge in Transition
,
1994
.
[15]
David E. Brown.
Facilitating conceptual change using analogies and explanatory models
,
1994
.
[16]
M. Chi,et al.
From things to processes: A theory of conceptual change for learning science concepts
,
1994
.
[17]
S. Vosniadou.
Capturing and modeling the process of conceptual change.
,
1994
.
[18]
R. Baillargeon.
A model of physical reasoning in infancy
,
1995
.
[19]
Nicola Fanizzi,et al.
Conceptual Change in Learning Naive Physics: The Computational Model as a Theory Revision Process
,
1999,
AI*IA.
[20]
Kenneth D. Forbus,et al.
SEQL: Category learning as progressive abstraction using structure mapping
,
2000
.
[21]
Jeffrey M. Zacks,et al.
Perceiving, remembering, and communicating structure in events.
,
2001,
Journal of experimental psychology. General.
[22]
Kenneth D. Forbus,et al.
Transforming between Propositions and Features: Bridging the Gap
,
2005,
AAAI.
[23]
Douglas H. Fisher,et al.
Knowledge Acquisition Via Incremental Conceptual Clustering
,
1987,
Machine Learning.
[24]
Johanna D. Moore,et al.
Proceedings of the 28th Annual Conference of the Cognitive Science Society
,
2005
.
[25]
Kenneth D. Forbus,et al.
SpaceCase: A Model of Spatial Preposition Use
,
2005
.
[26]
Kenneth D. Forbus,et al.
Companion Cognitive Systems: Design Goals and Some Lessons Learned
,
2008,
AAAI Fall Symposium: Naturally-Inspired Artificial Intelligence.
[27]
Scott E Friedman.
Learning Causal Models via Progressive Alignment & Qualitative Modeling : A Simulation
,
2008
.
[28]
E. Bard,et al.
Proceedings of CogSci 2008
,
2008
.
[29]
CogSketch
,
2008,
AAAI.
[30]
Kenneth D. Forbus,et al.
EA NLU: Practical Language Understanding for Cognitive Modeling
,
2009,
FLAIRS.