PaperlayerView exact revision on arXiv

Source-linked AI summary

A Personalized and Adaptive Intelligent Tutoring System for Psychomotor Learning

Madeleine Shuhn-Tsuan Yuh, Neera Jain

arXiv:2610.33467v1eess.SYcs.HCcs.SY

TL;DR

Psychomotor ITSs often rely on learner models established before instruction or updated only once per module, leaving open how to personalize tutoring from ongoing assessments. This paper adapts objective weights using recent learner states and evaluates the approach in a 40-person quadrotor study, finding distinct prioritization patterns and an association between weight changes and higher final performance.

  • Problem

    Existing psychomotor ITSs often use learner models built before instruction or updated once per module, leaving open how ongoing assessments should inform personalization.

    Source: §I · pp. 1–2

  • Method

    An actor–critic-inspired algorithm dynamically reprioritizes tutoring objectives through reward weights within a cognitive-control framework for psychomotor ITSs.

    Source: §I · p. 2, §VII · p. 10

  • Results

    Learners whose reward-weight vectors shifted enough to move clusters generally achieved higher final performance, except for the workload-calibration group.

    Source: §VII · p. 10

  • Takeaways & Limitations

    The work provides a computational framework for personalization during psychomotor learning and reports quantitative benefits of personalization.

    Source: §VII · p. 11

    The study included 40 participants, and some cluster-defined groups were small, with the smallest containing seven participants.

    Source: §VII · p. 11

Abstract

from arXiv · show

Intelligent tutoring systems (ITSs) have advanced learner motivation and personalization in cognitive domains, but extending these gains to psychomotor tasks remains challenging. Existing psychomotor ITSs rely on learner models built before instruction or updated once per module, leaving open how to use real-time assessments of learner cognitive state for finer-grained personalization. Psychomotor learning involves multiple tutoring objectives, including calibrating self-confidence to performance, calibrating workload per the Yerkes-Dodson Law, and driving skill progression toward higher learning stages. How to balance these objectives for learners, and adjust that balance over time, remains open. This study proposes and evaluates an online personalization algorithm that continuously reprioritizes tutoring objectives to match learners' evolving needs. Rather than weighting objectives equally, the actor-critic-inspired algorithm dynamically updates reward weights using recent measures of self-confidence, workload, and performance. These weights inform a multi-input, multi-output cognitive control framework that synthesizes learner-specific tutoring policies via model-based planning. The approach was evaluated in a human-subject study (N = 40) with a 2D quadrotor landing task. Clustering of the learned reward weights revealed distinct prioritization patterns across learners, indicating the algorithm captures heterogeneity in learner needs. Learners whose weights shifted more substantially tended to reach higher final performance, and those whose weights consistently emphasized workload calibration achieved the strongest outcomes. These findings advance online learner modeling and control-based personalization in psychomotor ITSs, showing how adaptive tutoring can better support learning by aligning instructional decisions with learners' evolving cognitive and performance states.

I. INTRODUCTION

Psychomotor tutoring systems need finer-grained personalization: existing systems often build learner models before instruction or update them only once per module. This paper proposes an algorithm that reprioritizes tutoring objectives during learning, with background on the quadrotor task and its online learning measures.

I. INTRODUCTION

  • Existing psychomotor ITSs often rely on learner models built before instruction or updated once per module, leaving open how to use ongoing assessments for finer-grained personalization.The paper proposes more frequent online updates to learner models and tutoring policies as task complexity and learner needs evolve.Source: §I · p. 2
  • The proposed actor–critic-inspired algorithm updates tutoring-objective priorities during learning using learner performance and cognitive states.Reward weights represent objective priorities; the paper reports distinct patterns across learners and statistical evidence that priorities correspond to learner needs.Source: §I · p. 2

A. Psychomotor Task: 2D quadrotor landing simulator

  • The study uses a 2D quadrotor landing simulator in which learners control thrust and roll, either manually or with automation assistance.Safe landings must satisfy speed and roll-angle constraints; crashes before landing count as unsuccessful.Source: A. Psychomotor Task: 2D quadrotor… · p. 2

B. Online Quantification of Learning

  • Learning is quantified after each trial with a discrete learning-stage metric and a continuous score suited to gradient-descent-based policy updates.These measures provide complementary characterizations of progression and performance.Source: B. Online Quantification of Learning · pp. 2–3

1) Discrete-valued Learning Stage Classification:

The section describes a discrete classifier for skill progression and a continuous score for online policy adaptation. The score combines landing-condition measures with similarity to proficient trajectories.

1) Discrete-valued Learning Stage Classification:

  • The classifier assigns quadrotor learners to the first four Dreyfus stages, from novice through competent, using trajectory data from task execution.Stage definitions distinguish feasible landing outcomes and trajectory characteristics; representative trajectory distributions underpin the classifier.Source: 1) Discrete-valued Learning Stage… · p. 3

2) Continuous-valued Performance Score:

  • The continuous score combines position, velocity, roll-angle, and proficient-trajectory similarity scores to quantify performance for online policy updates.A score of 750 out of 1000 indicates a successful landing that is not yet representative of a canonically proficient learner.Source: 2) Continuous-valued Performance Score · pp. 3–4

C. Human Learner Model

The learner model represents stochastic transitions in cognitive states and learning progression as an MDP. It supports tutoring-policy design by linking learner states and tutoring actions to rewards.

C. Human Learner Model

  • The MDP models stochastic transitions in self-confidence, workload, and learning-stage states to support tutoring-policy design.Tutoring objectives are represented through a reward function of learner states and actions.Source: C. Human Learner Model · p. 4
  • Self-confidence and workload are discretized into low, medium, and high states, while learning stages 3 and 4 are combined into one state.The self-report thresholds are the 33% and 66% quantiles; stages 3 and 4 are merged because their quadrotor trajectory behavior is similar.Source: C. Human Learner Model · p. 4
  • Tutoring actions encode whether automation is on or off and whether feedback is negative, neutral, or positive.State-transition models for the cognitive states and learning progression are trained independently under an independence assumption previously supported in.Source: C. Human Learner Model · p. 4

III. TUTORING POLICY DESIGN AND ADAPTIVE PERSONALIZATION

The tutoring policy combines three objectives in a weighted reward and uses model-based optimization to choose actions. Because learners’ needs can differ and change, the paper motivates adapting the weights over time.

A. Reward Function Design and Policy Training

  • The reward function covers confidence calibration, workload calibration under the Yerkes-Dodson law, and progression toward higher learning stages.The first two objectives are motivated as components of self-efficacy, while the third explicitly encourages skill progression.Source: A. Reward Function Design and Policy… · p. 4
  • Weights α1, α2, and α3 encode the relative priority of the three objectives, and Bellman optimality equations are used to compute the policy.For each state, the policy selects the action that maximizes the action-value function.Source: A. Reward Function Design and Policy… · p. 4

III. TUTORING POLICY DESIGN AND ADAPTIVE PERSONALIZATION

  • The paper motivates adapting reward weights because objective priorities may differ across learners and change over time.The preceding section introduces an algorithm that adjusts the weights based on individual performance over consecutive trials.Source: A. Reward Function Design and Policy… · p. 4, §III · p. 4

B. Reward Weight Adaptation Algorithm

The algorithm adapts tutoring-objective priorities from recent learner data, then uses those priorities to guide policy synthesis. Its objectives track confidence against performance, workload against performance, and progress toward consistent proficiency.

B. Reward Weight Adaptation Algorithm

  • Sparse, fixed-policy data and non-stationary learner behavior limit online reinforcement learning, while aggressive exploration may harm learning or user experience.The data also provide limited state–action coverage and short episodes, constraining long-horizon credit assignment and value estimation.Source: B. Reward Weight Adaptation Algorithm · pp. 4–5
  • The actor–critic-inspired approach uses recent observations to reweight tutoring objectives, then computes policies through model-based planning rather than online interaction.The paper describes this as a more stable and data-efficient approach than policy updates through online interaction.Source: B. Reward Weight Adaptation Algorithm · p. 5

1) Objective Function Definition:

  • Three objective functions measure confidence–performance alignment, inverse workload–performance alignment, and progress toward a performance threshold with consistency.Confidence and workload are self-reported; the performance objective combines a penalty for falling below the threshold with a variance penalty.Source: 1) Objective Function Definition · p. 5

2) Reward Weight Adaptation via Gradient Descent:

Recent learner data produce a target priority distribution that raises emphasis on objectives not being met. Gradient descent fits reward weights to that distribution while regularizing toward prior weights, after which model-based Q-iteration computes the tutoring policy.

2) Reward Weight Adaptation via Gradient Descent:

  • Unsatisfied objectives receive greater target reward weight, and larger objective losses produce larger target weights.The temperature parameter κ controls how strongly the least-satisfied objective dominates the target distribution.Source: 2) Reward Weight Adaptation via… · p. 5
  • Gradient descent fits weights to the target distribution while keeping them close to a prior, discouraging abrupt shifts in tutoring priorities.The prior regularization strength is controlled by β; the weights form a nonnegative convex combination that sums to one.Source: 2) Reward Weight Adaptation via… · p. 5
  • After adapting weights from recent learner data, the method constructs a weighted reward and applies model-based Q-iteration to obtain an action-value function and tutoring policy.The algorithm returns the learned weights, weighted reward, optimal action-value function, and policy.Source: 2) Reward Weight Adaptation via… · pp. 5–6

IV. HUMAN USER STUDY

The study evaluates personalized tutoring during a 2D quadrotor landing task, using phased tutoring and manual trials to assess learning. Forty participants completed the study.

IV. HUMAN USER STUDY

  • The study tests whether learned weights show distinct patterns, change with learner needs, and prioritize needed objectives during a 2D quadrotor landing task.The tutoring assistance is guided by the proposed reward-weight adaptation algorithm.Source: §IV · p. 6

A. Experiment Procedure

  • Participants first completed two manual trials without feedback, then 30 tutoring trials across three phases with randomized quadrotor starting positions.Each phase began with uniformly initialized reward weights; tutoring ran for seven trials, followed by two manual trials and a policy update.Source: A. Experiment Procedure · p. 6
  • After Phase 3, participants completed five manual trials without formative feedback to measure final manual performance.Each phase used an updated reward-weight vector to synthesize the policy for the next phase.Source: A. Experiment Procedure · p. 6

B. Participants

  • Forty participants aged 18–70 completed the study; the sample included 10 men, 29 women, and one non-binary participant.The study was approved by Purdue University’s Institutional Review Board.Source: B. Participants · p. 6

C. Apparatus

The analysis clusters reward-weight vectors to characterize tutoring priorities and their changes across phases, then compares learner outcomes across groups. It also defines manual-trial measures of performance and cognitive state.

C. Apparatus

  • The simulator used a joystick and throttle to control the quadrotor in a Python and Pygame implementation.The simulator was adapted from the platform described in.Source: C. Apparatus · p. 6

D. Data Collection

  • Final performance is the mean score over the last five manual trials, while change in performance compares those trials with the first two manual trials.Final confidence and workload are means of self-reports during the same five final manual trials.Source: D. Data Collection · p. 7

V. DATA PROCESSING AND STATISTICAL ANALYSIS

  • Of 40 participants, 9 consistently prioritized confidence calibration, 8 workload calibration, 7 learning-stage progression, and 16 shifted clusters.The cluster assignments were derived from reward-weight vectors learned after each phase.Source: §V · p. 7, D. Data Collection · p. 7
  • Silhouette analysis selected k=3 clusters while the elbow method suggested k=4, with the discrepancy indicating similar underlying structure.The authors prioritized silhouette-based selection; the resulting clusters represent confidence calibration, workload calibration, and learning-stage progression.Source: §V · p. 7, Figure 3 · p. 8
  • ANOVA found significant group differences in final performance and self-confidence, so post hoc comparisons were conducted for those outcomes.The analysis also tested change in performance and final workload across groups.Source: §V · p. 8, Table I · p. 8

VI. RESULTS AND DISCUSSION

Learners showed distinct priorities among tutoring objectives, and participants whose reward weights changed across phases often reached scores above the safe-landing threshold. The study could not determine whether those participants benefited compared with a static policy.

the tutoring objective weights across all learners?

  • The three-cluster solution identified priorities for self-confidence calibration, workload calibration, and learning-stage progression; the four-cluster solution added a balanced emphasis on confidence and workload.The three-cluster solution was selected for subsequent analysis as the best trade-off between cluster quality and model complexity.Source: the tutoring objective weights across… · p. 8, Figure 4 · p. 9

to the learner’s needs?

  • 814.39 (SD 170.96) was the change group’s mean final score across the last five trials; most in that group scored above 750.A score of 750 is required for a safe landing.Source: to the learner’s needs? · p. 8, Table II · p. 9
  • The between-subjects study could not establish whether participants with changing reward weights benefited compared with a static policy, despite no statistically different outcomes from other groups.Source: to the learner’s needs? · pp. 8–9

the tutoring objectives learners need?

The workload-calibration group had the highest and most consistent final performance, while other groups’ outcomes aligned with different tutoring-objective priorities. Reward-weight shifts and cluster patterns tracked learner differences; the study also notes small cluster groups and proposes further validation and model extensions.

the tutoring objectives learners need?

  • Group comparisons found no significant differences in change in performance or final workload, and generally no significant pairwise differences in final performance or confidence.The reported pairwise comparisons had a few exceptions.Source: the tutoring objectives learners need? · p. 9, Table III · p. 10
  • 651.66 mean final score in the learning-stage group was near the scoring midpoint, with mean confidence 55.29 and workload 52.00 calibrated to performance.The interpretation is that the group’s weights prioritized stage progression, its least-met objective.Source: the tutoring objectives learners need? · p. 9, Table IV · p. 10
  • 933.95/1000 mean final score (SD 35.24): the workload-calibration group had the highest mean and lowest performance variability across groups.Its workload was nevertheless elevated relative to its high performance, consistent with prioritizing workload calibration.Source: the tutoring objectives learners need? · pp. 9–10, Table IV · p. 10
  • The workload-calibration group’s mean final workload was 45.75/100 (SD 20.70), elevated relative to its high performance and consistent with prioritizing workload calibration.The group also reported the highest mean self-confidence and significantly higher confidence than the self-confidence and learning-stage groups.Source: the tutoring objectives learners need? · p. 10, Table III · p. 10, Table IV · p. 10
  • Despite a second-highest mean score of 793.27/1000, the self-confidence group averaged 46.67/100 confidence versus approximately 80.00/100 expected for calibrated performance.Its weights were interpreted as prioritizing confidence calibration, the group’s least-calibrated objective.Source: the tutoring objectives learners need? · p. 10, Table IV · p. 10

VII. CONCLUSION

  • The proposed framework dynamically reprioritizes confidence calibration, workload calibration, and learning-stage progression; future work includes modeling cognitive-state interdependencies and individualizing reward-weight update timescales.The authors also propose a more extensive user study.Source: §VII · pp. 10–11
  • Learners whose reward-weight vectors shifted enough to change clusters generally achieved higher final performance than learners who stayed in the same cluster, except in the workload-calibration group.The study used k-means clustering to identify distinct tutoring-objective weight patterns.Source: §VII · p. 10
  • With 40 participants, cluster sizes could not be controlled and some groups were small, with the smallest containing seven participants.The authors identify a larger user study as a way to further validate efficacy.Source: §VII · p. 11

APPENDIX

The appendix details the learner-model estimation and validation procedures and specifies reward functions for confidence calibration, workload calibration, and learning-stage progression.

APPENDIX

  • Learner-model parameters were estimated by optimizing initial and transition probabilities to maximize the likelihood of observed state-action sequences, using a genetic algorithm.The algorithm was available in MATLAB’s Optimization Toolbox.Source: APPENDIX · p. 12
  • Five-fold cross-validation trained and tested the model across all folds; the selected model had the highest combined training and test log-likelihood.Four folds trained the model while the remaining fold served as test data in each run.Source: APPENDIX · p. 12
  • Each three-state, six-action MDP model represents transitions in a 3 × 3 × 6 array, covering 54 possible state transitions.Table V illustrates self-confidence transitions under positive feedback with assistance off and on.Source: APPENDIX · p. 12, Table V · p. 12
  • The self-confidence reward combines equally weighted assistance and feedback-tone terms that reward actions matching reported confidence to learning stage.For severe overconfidence at stage 1, negatively toned feedback earns 0.60 versus 0.20 for positive tone.Source: APPENDIX · p. 12, Table VI · p. 12, Table VII · p. 12
  • The workload reward favors workload appropriate to learning stage, including assistance when high workload suggests insufficient mental capacity for unaided performance.At stages 3 and 4, low workload with no assistance earns the maximum reward of 1.0; at stage 1, low workload also earns 1.0 as appropriately challenging workload is expected.Source: APPENDIX · pp. 12–13, Table VIII · p. 13
  • The stage-progression reward favors stages 3 and 4 without automation, but favors stage 1 with automation, with rewards declining across later stages.Only the assistance action enters this reward because it shapes performance within each trial.Source: APPENDIX · p. 13, Table IX · p. 13
  • All reward-function components are designed to remain within [0, 1].Source: APPENDIX · p. 13
  • Table X reports cluster assignments in phases 1 and 2 and mean score, self-reported confidence, and workload for participants whose cluster changed.The measures cover trials 33–37.Source: Table X · p. 13, APPENDIX · p. 13
Loading 2610.33467v1…