Abstract
This paper outlines a framework for developing rational agents capable of functioning in complex, uncertain environments where time and computational power are strictly limited. By combining Active Logic for time-aware deductive reasoning with Inductive Logic Programming (ILP) for learning from experience, the proposed architecture allows agents to navigate intractable search spaces through the execution of partial plans and continuous self-improvement.
1. Introduction
In classical artificial intelligence, agents are typically designed to find a complete path from an initial state to a goal before initiating any action. However, in real-world scenarios or complex games, the search space is often too vast for exhaustive computation, and the agent’s knowledge of the environment may be incomplete or erroneous. This paper addresses the challenge of resource-bounded reasoning: determining when an agent should stop "thinking" and start "doing." The goal is to create an agent that can formulate effective partial plans and learn from its mistakes to improve future performance.
2. Agent Architecture
The proposed framework consists of three interacting modules designed to balance deliberation and execution.
2.1 The Deductor
The Deductor utilizes First-Order Logic and Situation Calculus to generate potential plans. A key innovation is the use of Active Logic (AL). Unlike standard logic, Active Logic is "time-aware," tracking the passage of time during the reasoning process itself. This allows the agent to handle contradictions—such as a belief being invalidated by a new observation—without the system crashing, as it can reason about the timing of its beliefs.
2.2 The Actor
The Actor serves as the system's manager. It monitors the Deductor’s progress, evaluates the generated partial plans, and determines the optimal moment to initiate action. This module is responsible for the transition from internal deliberation to external environmental interaction.
2.3 The Learner
Following the conclusion of a task or game episode, the Learner module analyzes the "game history" using Inductive Logic Programming (ILP). By generalizing rules from past successes and failures, the Learner helps the agent refine its plan evaluation criteria, ensuring that the same errors are not repeated in subsequent episodes.
3. Key Reasoning Strategies
The architecture introduces several novel concepts for managing limited resources:
- Information-Providing Plans: Rather than acting solely to reach a physical goal, the agent can execute plans designed to acquire information. These actions improve the "epistemic quality" of the agent's state, allowing it to prune future search branches that are inconsistent with new observations.
- Active Logic for Real-Time Thinking: By reasoning about its own reasoning time, the agent can adjust its strategy dynamically if a solution is taking too long to compute.
- Learning from Defeat: The authors propose using ILP to treat plans leading to failure as "positive examples" of what to avoid. This negative-constraint learning helps the agent quickly identify and bypass dangerous decision paths.
4. Evaluation Framework
The paper identifies the General Game Playing (GGP) competition as the primary testbed for this architecture. In GGP, agents are provided with the rules of an unknown game and must begin playing effectively immediately. This environment necessitates an agent that can learn and act under significant time pressure and uncertainty.
While the paper focuses on the conceptual framework, it highlights the novel application of ILP for plan evaluation rather than simple classification. This approach addresses the exploration/exploitation dilemma by using empirical data to guide logical deduction.
5. Conclusion and Future Work
The proposed architecture bridges the gap between high-level logical reasoning and empirical learning. While the current work is primarily structural, future research aims to extend the system to discover "subgoals" and "landmarks"—intermediate stepping stones to a final goal. Additionally, the authors intend to use the learning module to automatically correct errors in the agent's initial domain knowledge, further enhancing its autonomy in unknown environments.
Summary updated