Synthetic Expertise

In this chapter, our Model of Expertise is introduced. The Model of Expertise is implementation independent so can describe both biological and artificial systems. In Chapter 5, basic requirements of an expert are discussed, In Chapter 6, a review of cognitive architectures is presented. Chapters 5 and 6 form the basis of the expertise model introduced in this chapter.

Requirements for Expertise

Summarizing Chapters 5 and 6, an expert must possess general knowledge, problem-solving skills, and knowledge about how to perform tasks. An expert must have generic problem-solving and task knowledge, applicable to any situation, and also domain-specific problem-solving and task knowledge. Knowledge must be general in nature (knowledge about things), specific in nature (deep domain knowledge), and also common sense in nature. An expert must maintain episodic knowledge as well (knowledge about past experiences). As an evolving agent, an expert must also maintain a collection of models, both generic and domain specific, goals, and utility values. As an intelligent agent, an expert must be able to perceive the environment and recognize a set of environmental configurations. Finally, an expert must have a set of actions able to be performed to effect a change in the environment.

An expert must be able to extract from its knowledge information about what to do in a given situation. The agent's domain-specific models, domain-specific knowledge, and episodic knowledge facilitate recognition of situations. Extraction of problem-solving and task knowledge relevant to the situation completes the "extraction" skills of an expert. An expert must be able to recall, apply, evaluate, understand, analyze, and create with respect to its domain of discourse. We also expect an expert to be able to teach others about the domain.

Finally, learning is inherent in all aspects of an expert. As an expert experiences new situations, knowledge of all types is extracted and used to enhance the agent's knowledge stores. The agent also continually updates its collection of models as well as its problem-solving and task knowledge. As an evolving agent, an expert continually evaluates and updates its goals and utility values.

The Knowledge Level Description of Expertise

The knowledge an expert must possess is described at the Knowledge Level as shown in Fig. 7-1. Here we combine types of knowledge from the formal models of intelligent agents together with the kinds of knowledge identified by Simon, Steels, and Gobet. Because this representation is at the Knowledge Level, implementation details about how the knowledge stores are realized are not specified nor implied. Refer to Chapters 5 and 6 for details.

An expert is in general an intelligent agent perceiving the environment, reasoning about those perceptions using its internal knowledge, and performing actions thereby causing changes in the environment. Because of limitations in its sensory systems, an expert perceives only a subset

of the possible states of the environment, T. The expert also has a set of actions, A, it can perform to change the environment. Experts are goal- driven and utility-driven evolving agents, as described in Chapter 6, where G represents the set of goals and U represents the set of utility values.

In addition to deep domain knowledge K_D (knowledge about the domain) experts possess general background knowledge К (generic knowledge about things), common-sense knowledge K_c, and episodic knowledge K_E (knowledge from and about experiences). A model, similar to Gobet's templates and Minksy's frames is an internal representation allowing the expert to classify its perceptions and recognize/differentiate situations. For example, an expert plumber would have an idea of what a leaky faucet looks, sounds, and acts like based on experience. This mental model of a leaky faucet allows the plumber to quickly recognize a leaky faucet when encountered. Some models are domain-specific, M_D, and other models are generic, M. In humans, the collection and depth of models is attained from years of experience. As models are learned from experience, creating and maintaining M_D requires K_E and K_D as a minimum but may also involve other knowledge stores.

As Simon and Steels identified, an expert must know how to solve problems in a generic sense, P, and how to solve problems with domain- specific methods, P_D. These represent the problem-solving knowledge of the expert. In addition, experts must also know how to perform generic tasks, L, and domain-specific tasks, L_D.

An expert is always learning therefore all knowledge types, models, task models, and problem-solving methods are continually changing. Experts can learn both generic and domain-specific forms of new knowledge and also refine knowledge already stored. An expert uses its knowledge to reason about what it perceives, infer and deduce causes and relationships, and consider the effect of possible actions.

However, the Knowledge Level description of expertise is incomplete because it does not include the skills an expert must possess. For this purpose, we introduce the Expertise Level as shown in Fig. 7-2.

The Expertise Level

The knowledge and skills an expert must possess are described in Chapter 5. An expert's knowledge is described above at the Knowledge Level as shown in Fig. 7-1. However, a model of expertise must accommodate both expert knowledge and expert skills. We seek a way to represent skills in a way not requiring us to worry about the knowledge required to perform those skills. Therefore, we introduce a new level called the Expertise Level lying above the Knowledge Level as shown in Fig. 7-2. The Expertise Level represents skills an expert must possess. At

Fig. 7-2: Tine expertise level.

the Expertise Level, we talk about what an expert does—the skills—and not worry about the details of the knowledge required to perform these skills. Therefore, the medium of the Expertise Level is skills.

The Expertise Level Description of Expertise

Figure 7-3 shows the skills needed by an expert—the Expertise Level description of an expert. Being an intelligent agent operating in an environment necessitates the perceive and act skills. Being a goal-driven intelligent agent, the alter skill allows the expert to change its goals over time. An expert acquires general knowledge, domain-specific knowledge, problem-solving knowledge, and task knowledge over time so the learn skill is necessary. Following studies of experts from Simon, Steels, and Gobet, among others, the extract skill is the way an expert matches its perceptions about a current situation with its knowledge and retrieves relevant chunks. We add the teach skill because we believe any expert should be able to teach someone else about their domain of discourse. To these skills, we add the six skills identified in Bloom's Taxonomy: recall, understand, apply, analyze, evaluate, and create. Note the assess function, allowing an intelligent agent to change utility values, is subsumed by the evaluate skill.

The result is twelve fundamental skills needed by an expert: recall, apply, evaluate, understand, analyze, create, extract, teach, learn, alter, perceive, and act. We call these the fundamental skills because any other skill an

Fig. 7-3: Expertise-level description of expertise.

expert may exhibit is a combination of several of these fundamental skills. We talk about higher-level composite skills in a later section.

As described in Chapter 5, the extract skill represents the expert's ability to match current perceptions, T, with its episodic and domain knowledge as well as with its models (K_E/ K_D, M, M_D) and retrieve chunks of problem-solving, task, episodic, and domain knowledge pertinent to the situation (P, P₀, L, L₀, K_r/ K_D). The expert understands and analyzes what it perceives using common sense and general knowledge (К, K_c) along with what it recalls and extracts from its knowledge stores. The expert can then evaluate the effect of possible actions considering the context of its goals (G) and the weight of its utility values (IT). Ultimately, the expert selects an action, A, and acts thereby changing the environment and starting the cycle over again. In parallel with this perceive-reason-act cycle, the expert continually learns new knowledge, alters its goals if necessary, and evaluates its utility values considering its recent experience.

The Model of Expertise

Figure 7-4 is a combination of Figs. 7-1 and 7-3, the Knowledge Level and Expertise Level description of an expert, yielding the formal Model of Expertise.

Fig. 7-4: Formal model of expertise.

The Soar Architecture for Expertise

Having identified the requisite knowledge and skills of an expert, we can now situate those into a Soar-based cognitive architecture as shown in Fig. 7-5. For this, we superimpose our formal model onto the basic Soar architecture of an evolving agent.

Fig. 7-5: The Soar model of expertise.