Introduction
Engineering knowledge for artificial intelligence systems is notoriously expensive and time-consuming. While "transfer learning" aims to mitigate this by reusing knowledge across tasks, most systems struggle with "far transfer." This occurs when the source and target tasks share an underlying logical structure but bear little surface similarity—for instance, they might use different names for objects, different sets of actions, or different relational symbols.
This paper introduces a "representation mapping" mechanism for the ICARUS cognitive architecture, enabling far transfer. By treating transfer as a goal-driven analogy, the system can translate abstract hierarchical skills from a source domain (like building a bridge) to a target domain (like tying barrels together), significantly accelerating problem-solving in unfamiliar environments.
The Representational Divide
The core challenge addressed by the authors is the "representational divide": how can an agent that knows how to "nail logs" to cross a river understand that it can "tie barrels" to achieve a similar goal in a different game? Traditional architectures often fail here because their knowledge is tied to specific symbols. The challenge is to integrate a general mechanism for analogy into a functional cognitive architecture that handles perception, reasoning, and action simultaneously.
Methods: The Representation Mapper
The authors extend ICARUS, a cognitive architecture that represents knowledge as hierarchical skills and concepts. The core innovation is a new module called the Representation Mapper. The process works in three main stages:
- Source Analysis: The agent solves a source problem and generates a "solution trace." It then analyzes this trace to explain why the solution worked based on its internal theory of the source domain.
- Mapping (The Analogy Engine): The mapper takes the source explanation and attempts to replicate it using the target domain's vocabulary. It looks for a one-to-one correspondence between source and target predicates and constants. This search is guided by structural, semantic, and pragmatic constraints.
- Translation and Reuse: Once a map is found, a Representation Translator converts the source skills into the target language. These "new" skills are then injected into the agent’s memory to guide behavior in the target task.
Experimental Evaluation
The researchers evaluated their system using the General Game Playing (GGP) framework. They developed 11 challenge scenarios categorized into three types:
- Homeomorphic: Tasks with one-to-one symbol correspondence but some "extra" irrelevant objects in the source.
- Reformulation: Isomorphic problems where all symbols are systematically replaced to obscure the relationship.
- Cross-domain: Disparate games that share high-level logic, such as 2D grid navigation and tool use.
Key Results and Insights
The authors performed a lesion study, comparing the full system against a version where the representation mapping was disabled.
- Success Rate: The mapping-enabled agent achieved positive transfer in 9 out of 11 cases. Without mapping, transfer was essentially zero.
- Performance Boost: In the most successful case (CrossDomain-4), the agent solved the target problem more than 10 times faster than it would have through raw exploration.
- Goal-Driven Abstraction: By focusing on the explanation of a goal's success, the system automatically ignores "distractor" objects. If a hammer was used in the source but isn't needed for the target logic, the mapper simply drops it.
Limitations and Future Work
Currently, the system assumes a relatively clean one-to-one isomorphism. Real-world "far transfer" might involve many-to-one mappings or more complex structural shifts. Additionally, finding the "best" map can be computationally expensive. Future versions should interleave mapping with performance, allowing the agent to "fix" its analogies as it learns more about the target domain and develops a long-term episodic memory.
Summary updated