Abstract
Machine learning approaches to knowledge compilation seek to improve the performance of problem-solvers by storing solutions to previously solved problems in an efficient, generalized form. The problem-solver retrieves these learned solutions in appropriate later situations to obtain results more efficiently. However, by relying on its learned knowledge to provide a solution, the problem-solver may miss an alternative solution of higher quality—one that could have been generated using the original (non-learned) problem-solving knowledge. This phenomenon is referred to as the masking effect of learning.
In this paper, we examine a sequence of possible solutions for the masking effect. Each solution refines and builds on the previous one. The final solution is based on cascaded filters. When learned knowledge is retrieved, these filters alert the system about the inappropriateness of this knowledge so that the system can then derive a better alternative solution. We analyze conditions under which this solution will perform better than the others, and present experimental data supportive of the analysis. This investigation is based on a simulated robot domain called Groundworld.
Introduction: The "Mental Rut" of AI
In machine learning and automated planning, "knowledge compilation" (such as chunking or Explanation-Based Learning) is a standard technique to accelerate problem-solving. By caching solutions to past problems, an agent can simply retrieve a rule instead of reasoning from first principles.
However, this efficiency comes with a hidden cost: the masking effect. Similar to the psychological concept of Einstellung (a mental set), an agent may become "blinded" by its own learned knowledge. Because a learned rule is a generalization, it might apply to a new situation where it technically achieves the goal but is no longer the optimal choice. The learned knowledge "masks" the superior, original problem-solving knowledge that would have found a better path if the agent hadn't relied on its memory.
The Proposed Solution: Cascaded Filters
Eliminating the masking effect at "generation time" (when the rule is first learned) is often computationally intractable in dynamic environments. Instead, the authors propose a retrieval-time strategy using cascaded filters. This approach involves a two-stage verification process whenever a learned rule is retrieved:
- Stage 1: Approximate Filter: A computationally inexpensive check (e.g., "Is there an obstacle in the expected path?") to flag if the learned plan might be low quality.
- Stage 2: Exact Filter: If the first stage flags a potential issue, a more detailed, albeit slower, verification (e.g., a brief simulation) confirms if the plan is actually suboptimal.
If both filters agree the plan is poor, the system discards the learned rule and replans from scratch. This exploits the fundamental asymmetry that verifying a solution is often much cheaper than deriving one.
Experimental Domain: Groundworld
The strategies were evaluated in Groundworld, a continuous 2D simulated robot domain. It features walls that obstruct both vision and movement, and two types of agents: pursuit agents and evasion agents. The evasion agent's goal is to reach a destination while avoiding capture by hiding behind walls.
Key Insights and Results
The study implemented these agents within the Soar architecture. Key findings include:
- Efficiency Gains: In path-planning tasks, the RT-cascade approach provided speedups of 5.2x to 11.3x compared to systems that blindly replanned whenever any change was sensed.
- Minimal Overhead: The cost of the extra verification step was negligible (1.02x to 1.04x slowdown) when the filters correctly identified a bad plan.
- Robustness: The system remained beneficial even with simple, "imperfect" approximate filters, as the second-stage exact filter served as a safety net against unnecessary replanning.
Conclusion
The masking effect represents a specific type of overgenerality where a rule is logically "correct" but practically "wrong" due to inefficiency. By using cascaded filters at retrieval time, AI systems can maintain the speed benefits of learned knowledge without falling into the "mental ruts" that lead to suboptimal performance in dynamic worlds. Future work will focus on automatically generating these filters from the assumptions made during the learning process.
Summary updated