Generalizability of Goal Recognition Models in NarrativeCentered Learning

Generalizability of Goal Recognition Models in Narrative-Centered Learning Environments Alok Baikadi Bradford Mott Jonathan Rowe, James Lester North Carolina State University 1

Goal Recognition in Narrative-Centered Learning Environments § Task: Identify the specific objective that the player is attempting to achieve § Goal recognition models enable the following: • Preemptively augmenting narrative experiences • Assessing problem solving in narrative-centered learning environments • Iteratively refining learning environment 2

Generalization of Goal Recognition § Goal Recognition is typically very domain dependent • Plan libraries • Many domain-independent techniques are only evaluated on one domain § Research Question: Can a domain-specific goal recognition model be applied in a principled way to a new domain and achieve similar results? 3

Related Work § Goal recognition is a restricted form of plan recognition (Carberry 2001; Kautz & Allen, 1986; § Investigated widely in numerous domains (Blaylock & Allen, 2003; Charniak & Goldman, 1993; Lesh, Rich § IO-HMM approach for recognizing high-level goals in simple action-adventure game (Gold, 2010) § PHATT-based approach for behavior recognition in real-time strategy game (Kabanza, § N-gram and Bayesian network approaches for goal recognition to support dynamic narrative planning (Mott, Lee & Lester, 2006) § MLN-based approaches (Singla and Mooney, 2011 ; Ha et al. , 2011 ; Sadilek and Kautz, 2012) Singla & Mooney, 2011) & Sidner, 1999) Bellefeuille & Bisson, 2010) 4

Outline § Goal Recognition Approach § Goal Recognition Corpora § Evaluation & Discussion § Conclusions and Future Work 5

Markov Logic Networks (MLNs) § Statistical relational learning • Combines first-order relational reasoning with statistical learning • Input: A set of first-order predicate calculus formulae, along with weights • Formulae can be expanded into a Markov Random Field for learning and inference § The joint probability distribution is defined as: § Toolkit: Markov The. Beast (Riedel, 2008) 6

Representation Predicate Interpretation action(t, a) Action a happens at time t argument(t, a) Argument a observed at time t location(t, l) Player is at location l at time t state(t, s) The narrative is in state s at time t goal(t, g) Player is pursuing goal g at time t 7

Context in Goal Recognition § Actions are not always independent § Traditional goal recognition formulation allows for all previous observations • Can lead to sparsity issues § Solution: Look for key events in the history that provide insight to the player’s context § Use the structure of the narrative to provide the context 8

Discovery Events § Task progress is represented by a sequence of discovery events § Partial Answers to Central Questions are clues towards the solution § Provides a context for goal recognition: What has the user discovered? 9

Discovery Event Formulae § Milestone formulae recognize which discovery events have already occurred § Uses a cardinality constraint to capture existence 10

Outline § Goal Recognition Approach § Goal Recognition Corpora § Evaluation & Discussion § Conclusions and Future Work 11

CRYSTAL ISLAND: OUTBREAK § 8 th grade microbiology § Valve Software’s Source engine § Science mystery § Goal: Identify source and treatment of outbreak 12

CRYSTAL ISLAND: Introduction 1. Student plays the role of a new visitor to the island. 2. Student discovers that several team members have fallen sick. 13

CRYSTAL ISLAND: Gathering Information 3. Student gathers clues from sick team members. 4. Student asks the camp’s pathogen experts about microbiology concepts. 14

CRYSTAL ISLAND: Gathering Information 5. Student views microbiologythemed posters. 6. Student reads books about different types of pathogens. 15

CRYSTAL ISLAND: Hypothesis Testing 7. Student conducts tests using laboratory equipment. 8. Student interacts with the lab technician to view microscopic images of pathogens. 16

CRYSTAL ISLAND: Reporting Findings 9. Student presents findings and recommended treatment to camp nurse. 17

Corpus Collection § Eighth grade class from public middle school § 153 participants § No prior experience with CRYSTAL ISLAND § Played game for 1 hour, or until they were finished § 7 goals available to students (Ha et al. , 2011) 18

CRYSTAL ISLAND: UNCHARTED DISCOVERY Upper Elementary Science Subject § 5 th grade science § Standards aligned Content § Landforms § Maps, models & navigation Story § Adventurous adolescent § Shipwrecked crew § Complete quests to explore island 19

CRYSTAL ISLAND: UNCHARTED DISCOVERY Video 20

Corpus Collection § Onsite at 8 schools § 831 fifth grade students § 62% Caucasian, 14% African American, 8% Asian, 16% Other § Teacher-driven implementation in classrooms § 6 one hour sessions over 4 weeks § 12 goals available during the first 2 weeks 21

Goal Extraction Procedure § Goal-achieving actions were identified § Actions between previous goal and current goal were labeled with current goal § Goal-achieving actions were removed 22

Outline § Goal Recognition Approach § Goal Recognition Corpora § Evaluation & Discussion § Conclusions and Future Work 23

Empirical Evaluation § State of the Art Baseline: • Factored model (Ha et al. , 2011) • Uses MLNs to relate the current time step to the previous time step § Each model was evaluated using 10 -fold student-level cross -validation § Each model was evaluated according to three metrics: • Accuracy: Measured as F 1 score • Convergence rate: Percent of sequences which eventually predicted the correct goal • Convergence point: In sequences that converged, the percent of actions that had to be observed before a consistent prediction was made 24

Baseline Model 25

CRYSTAL ISLAND: OUTBREAK Discovery Events Model 26

CRYSTAL ISLAND: UNCHARTED DISCOVERY Discovery Events 27

EXPERIMENTAL RESULTS Crystal Island: Outbreak Model F 1 Convergence Rate Convergence Point Baseline 0. 488 30. 906 50. 865 Discovery Events 0. 546 50. 056 35. 862 Crystal Island: Uncharted Discovery Model F 1 Convergence Rate Convergence Point Baseline 0. 226 11. 915 87. 786 Discovery Events 0. 244 29. 973 79. 350 28

Outline § Goal Recognition Approach § Goal Recognition Corpora § Evaluation & Discussion § Conclusions and Future Work 29

Conclusions § Goal recognition models show considerable promise for enhancing the effectiveness of narrative-centered learning environments § Encoding narrative discovery events in Markov Logic is a natural approach for representing context for student actions in goal recognition § Experimental findings from two narrative-centered learning environments suggest that narrative discovery events enhance the accuracy and convergence of state -of-the-art MLN-based goal recognition models. 30

Future Work § Investigate combinations of discovery events • Some of the milestones may have provided more information than others • Use automated feature selection § Integrate goal recognition into a runtime environment • Can establish intuition for how accurate a model is necessary § Elicit feedback from player • Assumes goals achieved are intended • May cause some bias 31

Collaborators Research Staff Eleni Lobene Rob Taylor Affiliated Faculty Carol Brown (East Carolina University) Roger Conner (East Carolina University) Patrick Fitz. Gerald (Art + Design) Elizabeth Hodge (East Carolina University) James Minogue (Elementary Education) John Nietfeld (Educational Psychology) Marc Russo (Art + Design) Hiller Spires (Curriculum & Instruction) Eric Wiebe (STEM Education) Postdoc Eunyoung Ha Digital Art Staff Kirby Culbertson Sarah Hegler Karoon Mc. Dowell Graduate Students Julius Goth Joe Grafsgaard Eunyoung Ha Seung Lee Sam Leeman-Munk Wookhee Min Chris Mitchell Jennifer Sabourin Andy Smith Undergraduate Student Stephen Cossa Affiliated Post-Docs and Graduate Students (Art, Education, Psychology) Megan Hardy (Human Factors) Kristin Hoffman (Educational Psychology) Angela Meluso (Curriculum & Instruction) Lucy Shores (Educational Psychology) Sinky Zheng (Curriculum & Instruction) 32

Acknowledgments Support provided by the National Science Foundation under grant DRL-0822200. Additional support was provided by the Bill and Melinda Gates Foundation, the William and Flora Hewlett Foundation, EDUCAUSE, and the Social Sciences and Humanities Research Council of Canada. 33