Introduction
Standard Artificial Intelligence (AI) often falters when a problem requires tools or actions outside its pre-defined "vocabulary." Traditional planning and learning algorithms typically return a failure when encountering "off-nominal" or anomalous situations—for instance, a robot needing to use a heavy book as a hammer when no hammer is available. This paper defines Creative Problem Solving (CPS) as the process of expanding an agent's conceptual space to transform "un-achievable goals" into solvable ones. By providing a comprehensive framework, the authors aim to bridge the gap between traditional AI planning and the field of Computational Creativity.
The CPS Framework
The authors propose a four-stage framework for designing and analyzing CPS systems, unifying disparate research threads into a cohesive taxonomy.
- Problem Formulation: This stage determines if the task is approached via Planning (logical sequences) or Learning (typically Reinforcement Learning).
- Knowledge Representation: This describes how the agent stores information, categorized as Symbolic (logic-based), Non-symbolic (neural/sensorimotor), or Hybrid.
- Knowledge Manipulation: The "creative" core of the system, based on Margaret Boden’s three levels of creativity:
- Exploratory: Searching the "universal" conceptual space beyond current constraints.
- Combinational: Combining existing concepts in novel ways.
- Transformational: Altering the fundamental rules of the conceptual space so previously impossible solutions become apparent.
- Evaluation: Testing the new conceptual space through Simulation, Real-world robotics, or Benchmarks.
Key Insights and Meta-Analysis
As a survey and framework paper, the authors reviewed 51 existing works to validate their taxonomy. Several key insights emerged:
- Defining Creativity as Space Expansion: In an AI context, creativity is not merely "generating something new" but the functional expansion of what the agent perceives as possible. This creates a clear boundary between "hard planning" (searching within a space) and CPS (modifying the space itself).
- Cross-Pollination with Computational Creativity: The paper argues that AI has much to learn from Computational Creativity (CC)—a field usually focused on art or music—to solve functional engineering problems.
- Architectural Categorization: The meta-analysis identified four primary existing architectures that fit the CPS framework: CreaCogs, Robogyver, ICARUS, and DI-ARC.
- The Hybrid Gap: While hybrid representations (combining symbolic logic with neural learning) are theoretically powerful, they remain under-explored in creative contexts.
Research Gaps and Future Directions
The survey highlights a significant lack of standardized benchmarks in the CPS field, which hinders the ability to compare different creative AI approaches. Most current work is domain-specific and lacks "cross-domain generalizability"—for example, the ability to apply a "lever" concept learned in a cooking task to a woodworking problem.
Future research should focus on:
- Universal Metrics: Developing objective ways to measure "creativity" beyond simple task success.
- Lifelong CPS: Enabling agents to build a library of creative "hacks" over time rather than forgetting them after task completion.
- Algorithmic Efficiency: Implementing "Transformational" creativity remains computationally expensive and algorithmically difficult.
By formalizing the relationship between problem-solving and creativity, this framework provides a roadmap for developing agents capable of improvising in unstructured, real-world environments.
Summary updated