STAND: Self-Aware Precondition Induction for Interactive Task Learning

TL;DR

STAND tackles precondition induction in Interactive Task Learning by constructing a lattice of competing generalizations and leaf-specific constraints (G and S) and by reporting training progress through Cert-based certainty measures. It uses hierarchical shrinkage to stabilize probability estimates in small-data settings and enables active learning by highlighting edge cases with high disagreement. Across ITL domains and synthetic tasks, STAND achieves strong holdout accuracy, low error recurrence, and monotonic improvement, outperforming several baselines. The approach offers a data-efficient, interpretable, self-aware learning signal that can guide human instructors and potentially improve reliability in interactive robotic and software automation tasks.

Abstract

In interactive task learning (ITL), AI agents learn new capabilities from limited human instruction provided during task execution. STAND is a new method of data-efficient rule precondition induction specifically designed for these human-in-the-loop training scenarios. A key feature of STAND is its self-awareness of its own learning -- it can provide accurate metrics of training progress back to users. STAND beats popular methods like XGBoost, decision trees, random forests, and version spaces at small-data precondition induction tasks, and is highly accurate at estimating when its performance improves on holdout examples. In our evaluations, we find that STAND shows more monotonic improvement than other models with low rates of error recurrence. These features of STAND support a more consistent training experience, enabling human instructors to estimate when they are finished training and providing active-learning support by identifying trouble spots where more training is required.

Figures (14)

• Figure 1: Decision tree and STAND fit to the same input data. In STAND, multiple splits (filled grey circles) are expanded per node. STAND builds a general condition space (top-right) that is bounded below by a specific extension (bottom).
• Figure 2: Each model's holdout accuracy by example for Dice Adventure (via VAL), Fractions and Multi-column Addition (via AI2T), and synthetic data. STAND with hierarchical shrinkage matches or exceeds other models.
• Figure 3: For synthetic data, from left to right: the productive monotonicity by number of examples, true precision by estimated probability, and normalized active learning utility by example. Only showing prediction probability capable models.
• Figure 4: Average fit times in milliseconds for random forest (100 trees), XGBoost, decision tree, and STAND on fractions task. Neural net and VSSM are not shown (to maintain scale) since they take several seconds, which is not ideal for interactive training.
• Figure 5: Co-occurrence rate $P(X_i=v \;\land\; X_j=v)$ of categorical feature values for pairs of features $X_i$ and $X_j$. Matrix (a) shows co-occurrence just from uniform sampling. Matrix (b) has been structured by applying several conjuncts to the data at a rate of 80%.