TL;DR
This paper introduces an innovative information fusion algorithm (ITIIF) inspired by human cognitive processes and innovative thinking. By modeling stages like perception, memory, divergent/convergent thinking, and evaluation, ITIIF aims to achieve more effective information fusion and problem-solving with fewer evaluations, outperforming traditional intelligent algorithms in certain aspects.
What Problem Does This Paper Address?
Information fusion is crucial for integrating data from multiple sources into a unified knowledge mode to solve complex problems or generate new insights. While existing methods often rely on probability statistics or artificial intelligence (AI) techniques like neural networks and genetic algorithms, they tend to overlook the valuable knowledge generated during the problem-solving process itself. This limits the synergy between AI algorithms and knowledge, hindering their ability to fully leverage cognitive principles. The paper addresses this gap by proposing an algorithm that explicitly models human-like innovative thinking to enhance information fusion.
Methods
The ITIIF algorithm is built around five core modules:
- Information Sense and Perception: Gathers and processes raw data into a structured "knowledge cluster" (KC).
- Memory Storage: Organizes KCs into short-term memory (STM) and stores evolved "knowledge units" (KU) in long-term memory (LTM).
- Innovative Thinking: This is the core innovation, comprising:
- Divergent Thinking: Extracts KUs from LTM and assembles them into a "knowledge structure tree" (KST), representing various problem-solving paths.
- Convergent Thinking: Refines KST paths through focusing, modification, or innovation to arrive at an optimal solution.
- Evaluation System: Assesses the satisfaction degree of proposed solutions.
- Execution Module: Implements the chosen solution.
The algorithm emphasizes knowledge evolution within LTM, where KUs are continuously updated and improved. It also incorporates parameters for "creative style" (static vs. dynamic) to adapt its innovation strategy.
Datasets
The paper uses a dataset from "The Famous Knowledge Base for OpenCyc," containing 37,310 knowledge units and 6,531 correlations. This dataset is used to analyze the parameters of the ITIIF algorithm in the context of knowledge navigation path optimization. For performance comparison, the authors employ five standard test functions (f1-f5) commonly used in optimization research.
Key Insights and Ideas
- Bridging Cognitive Science and AI: The ITIIF algorithm is a significant attempt to translate abstract concepts from brain cognitive science and innovative thinking into concrete, operable computational steps for information fusion.
- Knowledge as a Core Component: Unlike many AI algorithms that treat knowledge as secondary, ITIIF places knowledge at its center, using it to drive the entire fusion and problem-solving process.
- Simulating Human-like Innovation: The explicit modeling of divergent and convergent thinking, along with adaptive creative styles, allows the algorithm to mimic human problem-solving and innovation, potentially leading to more robust solutions.
- Evolutionary Knowledge Base: The concept of knowledge evolution in LTM suggests a mechanism for continuous improvement and adaptation of the algorithm's knowledge over time.
- Reduced Evaluation Burden: The algorithm demonstrates the ability to find optimal solutions with fewer target evaluations, a key efficiency gain.
Key Results
- Parameter Influence: Analysis shows that parameters like memory extraction capacity (CLTM) and convergent dimension (ncon) significantly influence optimization performance. The "creative style" parameter (cre_style) also plays a role, with dynamic styles showing promise when target values are predictable.
- Performance Comparison: When tested against binary particle swarm and binary differential evolution algorithms on standard functions, ITIIF showed better optimization results in terms of accuracy and stability.
- Efficiency Gains: Compared to GA, EDAs, and ACO, ITIIF achieved better optimization effects, particularly in terms of solution accuracy and stability. While its average time was longer, this was attributed to its deeper use of knowledge, leading to significantly reduced target evaluation times (4.5% to 8% of other algorithms).
Limitations and Future Work
- Parameter Tuning: The paper acknowledges that parameter tuning (e.g., CLTM, ncon, cre_style) is crucial for optimal performance, and further research could explore more automated or adaptive tuning mechanisms.
- Theoretical Derivation: The authors suggest strengthening the theoretical derivation of the algorithm, particularly concerning the integration of innovative thinking.
- Broader Integration: Future work includes incorporating innovative thinking more deeply into the algorithm's framework and extensively integrating ITIIF with various traditional evolutionary algorithms to broaden its application areas.
- Specific Problem Instantiation: The evaluation functions within the algorithm (e.g., for knowledge radius) need to be instantiated when facing specific problems, indicating a need for domain-specific adaptation.
Summary updated