In artificial intelligence, the knowledge level is a level of abstraction for describing an agent's behavior purely in terms of its knowledge and goals, without reference to the underlying mechanisms that implement that knowledge. It was introduced by Allen Newell in 1982 in his paper "The Knowledge Level" (published in Artificial Intelligence, 18(1):87-127). At this level, an agent is characterized by what it knows and what it aims to achieve, and its actions are selected according to the principle of rationality: the agent chooses the action that it believes will best satisfy its goals given its knowledge.
The knowledge level is a logical abstraction that separates the content of an agent's beliefs and desires from the details of how those beliefs are stored or processed. It provides a way to rationalize an agent's behavior - to explain and predict what the agent will do based on its epistemic state and objectives, rather than on the specific algorithms or data structures it uses. This abstraction is useful for designing, analyzing, and comparing intelligent systems, as it allows reasoning about behavior at a high level of generality.
Relationship to the Symbol Level
Beneath the knowledge level lies the symbol level, which is system-oriented and includes the mechanisms an agent has available to operate. While the knowledge level is world-oriented, concerning the environment in which the agent operates, the symbol level encompasses the algorithms, data structures, and other implementation details that mechanize the agent's behavior. The knowledge level rationalizes behavior; the symbol level mechanizes it.
For example, in a computer program, the knowledge level consists of the information contained in its data structures that it uses to perform certain actions, such as facts about the world or rules for decision-making. The symbol level consists of the program's algorithms, the data structures themselves, and the physical or virtual hardware on which they run. The same knowledge-level description can be realized by many different symbol-level implementations.
Applications in AI Research
The knowledge level has influenced several areas of machine learning and generative AI. In knowledge-based agents, it provides a framework for specifying what an agent knows and what its goals are, independent of the inference engine or knowledge representation scheme used. This abstraction is particularly relevant in neural networks and large language models, where the internal representations are opaque but the knowledge-level behavior can be characterized by the model's outputs and objectives.
In deep learning, the knowledge level is often used to describe the functional competence of a model - what it has learned and what it can do - without detailing the millions of weights or the specific architecture. For instance, a transformer model trained on text can be described at the knowledge level as knowing facts about language and being able to generate coherent responses, even though its symbol-level implementation involves multi-head attention, positional encoding, and layer normalization.
Principle of Rationality and Goal-Directed Behavior
The principle of rationality is central to the knowledge level. It states that an agent will select an action that it believes will lead to the satisfaction of its goals, given its knowledge. This principle allows the knowledge level to be used predictively: if we know what an agent knows and what it wants, we can infer what it will do, without needing to know how it processes information internally.
This approach has been applied in reinforcement learning and curriculum learning, where agents are trained to maximize rewards (goals) based on observations (knowledge). It also underpins RLHF (reinforcement learning from human feedback) used in aligning OpenAI and Anthropic models, where the model's behavior is shaped by a reward signal that encodes human preferences.
Criticisms and Limitations
Some researchers have argued that the knowledge level is too abstract for practical use, as it assumes a complete and consistent specification of an agent's knowledge and goals, which is often unavailable in real-world systems. Others have noted that the distinction between knowledge and symbol levels can blur in systems that learn or adapt, where the knowledge itself is modified by symbol-level processes. Nevertheless, the knowledge level remains a foundational concept in cognitive science and AI, providing a clear vocabulary for discussing intelligent behavior.
See Also
- Knowledge level modeling
- Knowledge relativity
- knowledge-based agents
References
- Newell, A. (1982). The Knowledge Level. Artificial Intelligence, 18(1):87-127.
- Menzies, T. (1996). Applications of Abduction: Knowledge-Level Modeling. November 1996.