Introduction
Polyscheme is a cognitive architecture designed to integrate diverse AI algorithms—ranging from neural networks and logical solvers to robotic controllers—into a unified system. By decomposing these algorithms into a shared sequence of "attention fixations," Polyscheme enables disparate computational methods to collaborate in real-time. This approach addresses the traditional trade-offs between speed, generality, and robustness in artificial intelligence.
The Challenge of Integration
In AI research, no single algorithm is universally optimal. Search algorithms offer generality but are computationally expensive; reactive systems provide speed but lack planning capabilities; and structured representations like scripts are efficient but brittle. The primary obstacle to creating more capable systems is integration. These methods typically rely on incompatible data structures—such as graphs, rules, or tensors—making it difficult for them to communicate. Previous integration attempts often reduced one method to another, losing its unique advantages, or isolated them into modules with limited interaction.
The Polyscheme Architecture
Polyscheme overcomes these barriers through three foundational principles:
- Common Function Principle: Most AI algorithms can be broken down into basic functions: information storage, truth value estimation, forward inference, subgoaling, and representing alternate worlds.
- Multiple Implementation Principle: Modules called "Specialists" implement these functions using their native internal logic (e.g., a neural network specialist and a database specialist).
- Cognitive Self-Regulation: Algorithms are treated as strategies for addressing metacognitive issues like ignorance or conflict.
The core of Polyscheme is the Focus of Attention loop. A Focus Manager selects a specific proposition (e.g., "Is the robot behind the wall?"), and every specialist provides its opinion on that proposition's truth value. This sequence of fixations forms a Focus Trace.
Interleaving Traces for Hybrid Intelligence
The power of Polyscheme lies in its ability to interleave these traces. Instead of choosing between search and reactivity, the system can execute a hybrid algorithm. For instance, a Case-Based Reasoning module can "jump" a Search module closer to a solution by suggesting a likely state based on past experience.
As shown in the figure above, Polyscheme can combine the exhaustive exploration of search (A) with the efficiency of scripts (B). In the hybrid model (C), if a script encounters a gap (X), search specialists can temporarily take over to find a path to the goal (G), effectively bridging the gap between memory and planning.
Experimental Evaluation: Mobile Robot Tracking
The architecture was evaluated in a real-world scenario where a robot must track a target through occlusions and obstacles.
In this task, the Polyscheme-controlled robot used "path scripts" for rapid prediction but could switch to physical constraint reasoning when the target behaved unexpectedly. While traditional SAT solvers or grid-based planners often failed due to memory constraints, Polyscheme maintained efficiency by using structured scripts to prune the search space, only resorting to expensive search when necessary.
Key Results and Insights
- Attention as an Interlingua: Using the focus of attention as a bridge allows specialists to share information without needing to understand each other's internal data structures.
- Overcoming Brittleness: By combining memory-based scripts with active search, Polyscheme mitigates the brittleness of case-based reasoning. If a retrieved script contradicts the current environment, search specialists immediately intervene.
- Soundness in Hybridization: Implementation in database retrieval demonstrated that "fuzzy" methods like neural networks can be hybridized with formal logic while maintaining provable soundness and completeness.
Conclusion and Future Work
Polyscheme demonstrates a viable path toward truly integrated AI. However, several challenges remain. The current Focus Manager uses a simple queue; future iterations will require more sophisticated management to prioritize complex, long-term goals. Additionally, the system's intelligence remains dependent on the quality of its specialists and the metrics used for case retrieval. Future research will explore how this architecture scales to massive, open-world knowledge bases with numerous competing specialists.
Summary updated