Abstract
This paper presents a framework for AI systems to acquire skills through experiential learning. By transforming solutions from past experiences into plans for new problems, the system treats successful and failed transformations as training examples. These examples are then used to induce generalized plans, effectively compiling problem-solving experience into reusable, high-level skills.
Introduction
In the early 1980s, AI research excelled at concept acquisition—learning to categorize objects—but lagged in skill acquisition, the ability to solve complex, multi-step problems efficiently. Traditional planners like STRIPS or GPS often approached every new task from scratch using Means-Ends Analysis. Jaime Carbonell argued that this tabula rasa approach was un-human. Humans typically recall similar past situations and adapt their previous behaviors. Carbonell's work bridges the gap between "learning from examples" and "problem solving," positing that they are integrated processes.
Methodology: The Analogical Framework
Carbonell proposes a two-stage learning process centered on the concept of a Transformation Space.
Analogical Transformation (The T-Space)
Instead of searching through a "world space" of physical states, the system operates within a Transformation Space (T-space). Here, the search is conducted over plan modifications rather than raw actions.
- Initial State: A successful solution retrieved from a previous, similar problem.
- Operators (T-OPs): Edits such as "subgoal-preserving substitutions," "splicing in steps," or "deleting redundant steps."
- Goal: A modified solution sequence that satisfies the constraints of the new problem.
Inductive Generalization
After solving several related problems via analogy, the system uses these solutions as a training set for an induction engine.
- Positive Exemplars: Solutions that successfully achieved the goal in the real world.
- Negative Exemplars: Solutions that failed, particularly "near-misses."
The induction engine processes these to create a Generalized Plan—a parameterized script or macro-operator. This allows the system to solve future problems in the same class directly, bypassing the need for intensive analogical reasoning.
Key Insights
- Learning as an Inalienable Byproduct: Carbonell suggests that problem solving naturally provides the data necessary to tune future performance.
- T-Space Abstraction: Shifting the search from physical actions to plan edits significantly reduces the search space by starting from a "mostly correct" template.
- The Utility of Failure: Failed transformations serve as "near-miss" examples, which are vital for defining the boundaries of a generalized plan (e.g., identifying weight limits for a lifting plan).
- Self-Automated Learning: The system generates its own labeled training data by attempting analogies, reducing reliance on human-provided examples.
Results and Discussion
As a foundational conceptual paper, the focus is on the architectural framework. Carbonell noted that early implementations showed encouraging results, demonstrating that integrating learning with problem solving is computationally viable. Specifically, "near-miss" analysis helped refine the similarity metrics used to retrieve relevant experiences from memory.
However, several challenges remain. The system's efficiency is highly dependent on the quality of the initial similarity metric used for retrieval. As the library of past solutions grows, retrieval itself becomes a bottleneck—a problem that later spurred the field of Case-Based Reasoning (CBR). Future work aims to allow the system to learn from the planning process itself, rather than just the final sequence of actions.
Summary updated