Schema-based learning is a central theoretical approach in cognitive and educational psychology as well as in artificial intelligence. Schemas allow learners to reason about unfamiliar learning situations and interpret these situations in terms of their generalized knowledge. In cognitive and educational psychology, schema-based learning is grounded in capturing and using expert-generated schemas as frameworks for teaching and learning. Schemas can be learned to promote the acquisition of new scientific knowledge and skills. At the core of this approach is a generalization and abstraction method designed to extract and condense as much information as possible from a single example of successful task completion or problem solving.
In the field of artificial intelligence, schema-based learning is conceived as a generalized framework for the design of integrated adaptive autonomous agents aiming at the incorporation of general principles of adaptive organization and coherence maximization. Schema-based learning allows the development of increasingly complex patterns of interaction between the agent and its environment by confining statistical estimation to a narrow criterion.
Schema-based learning builds on the schema theory. The theoretical assumption of schema-based learning is that newly gained knowledge is assimilated into preexisting knowledge and organized to form schemas. However, it is not enough to simply have a collection of passive schemas: one must also know how to use them.
Anderson cites two types of situations involving schemas, namely: activation of preexisting schemas and construction of new schemas. However, Bransford (1984) points out that while it is possible to activate existing schemas with a given topic, this does not necessarily mean that a learner can use this activated knowledge to develop new knowledge and skills. Accordingly, Bransford stresses the importance of helping learners to “activate various preexisting ‘packets’ of knowledge” and to “reassemble and construct this knowledge into an integrated new schema” (Bransford 1984, p. 264). Bransford and colleagues (1981) also argue that the new learning situation should help learners to fully understand the “significance of new facts” as it helps them to reorganize the previously unrelated facts in a more meaningful way.
It is advantageous (efficient) to cope with new experiences on the basis of positive results from previous similar experiences.
Schema-based learning allows incremental development of complex cognitive structures through aggregation from a restricted stock of schemas (i.e., stable units of composition) to more complex interactive structures.
Schema-based learning provides the learner with instantiation, accretion, tuning, and reconstruction of knowledge. As a consequence, new schemas can be formed.
At the heart of this approach is a generalization and abstraction method which is designed to extract and condense as much information as possible from single examples of successful task completion or problem solving.
Internal-schema constraints connect component schemas together to form a complex schema. Technically, these are the assertions which support goal achievement by identifying parameters of instances of schematic forms with instances of prototypes.
Intra-schema constraints ensure that each component of a complex schema is complete. Technically, these are the immediate supporters of the “self” constraints which support the achievement of the goal.
Optional constraints are “forced” or implied by essential constraints. They need not be included but should be. They do not alter generality but improve efficiency. Technically, any assertion which has some justification depending purely on essential constraints is a member of this class. Extraneous constraints include most instantiation bindings and all implications based at least in part on extraneous constraints. Schema-based learning allows the development of increasingly complex patterns of interaction between the person (or an agent) and its environment by starting with a limited stock of simple schemas that allow efficient learning by confining statistical estimations of future events within the realm of a relatively narrow space.
The literature on schema-based learning theories describes three types of learning: accretion, tuning, and cognitive restructuring (Rumelhart and Norman 1978). Research conducted in the area of cognitive restructuring is generally categorized under the heading of conceptual change. A traditional method used to facilitate conceptual change is to provide the learner with examples that contradict their “naïve theories,” which is referred to as the anomalous data approach. Schema construction is a central element in schema-based learning. Learning is triggered when predictions made by anticipatory schemas do not match the observed results and the system does not reach the goal expectations. The dynamics of the system first try to reduce the error by tuning or accreting the current stock of schemas, and when this fails, a new schema must be constructed to remedy the error. In the long term, the reliability of the predictions will increase, making the system increasingly better at predicting the results of actions for a given context. In many cases, the agent will construct both a new schema and its corresponding predictive schema. Actually, most predictive/anticipatory schemas have to be learned by the agent through continuous interactions with the environment.
In the field of informatics and artificial intelligence, schema-based learning is a data-driven, constructivist approach used to discover probabilistic action models within environments that serve as a generalized framework for designing integrated adaptive autonomous agents and predicting their actions by incorporating general principles of adaptive organization and coherence maximization (Corbacho 1998). A schema is defined as an experience-based recurrent pattern of interaction with the environment, and coherence is a measure of the congruence between the result of an interaction with the environment and the expectations the agent has for that interaction.
Schema learning comprises two basic phases: discovery, in which context-free action/result schemas are found, and refinement, in which context is added to increase reliability.
Important Scientific Research and Open Questions
In cognitive and educational psychology, schema-based learning is grounded on the concept of equilibration in Piaget’s epistemology. According to Piaget, equilibration involves both assimilation and accommodation. During each stage of development, people conduct themselves with certain logical internal mental structures that allow them to make adequate sense of the world. When the newly learned information packet does not match with the learner’s existing internal mental structures (existing schema), equilibration helps the learner to make sense of the world around him by assimilating new information into pre-existing mental schemes and accommodating it when necessary through logical thinking. This internal mental process enables the learner to construct more sophisticated schemas, which in turn leads to cognitive development.
Pascual-Leone retained Piaget’s view of the human cognitive process as a highly dynamic and self-reflective system which passes through stages of stability and disequilibrium in the course of cognitive development. Pascual-Leone and Goodman (1979) distinguish several operators which are involved in the construction of new knowledge based on the assumption that the construction of new knowledge can be understood as a process of mental representation of environmental patterns. This process is strongly influenced by existing assimilative schemas. Accordingly, field forces compete with operators to determine which of the many schemas will be activated to regulate the processing of new information. During the course of processing, the activated schemas are enriched or restructured by internal operators. Pascual-Leone and Goodman distinguish between L-operators, which represent former learning experiences, A-operators, which represent emotional and affective side-effects, and B-operators, which represent stable personality traits. Finally, they add an M-operator, a moderator for using information processing capacity gained during development.
Ausubel’s theory of meaningful learning (1968) is also meant to help students activate their pre-existing knowledge so that it can be assimilated, tuned, and restructured into new schemas. Hence, Ausubel introduced the theory of “advance organizers” to support the learner’s pre-existing knowledge. Advance organizers can help learners to activate schemas more effectively and allow them to use their pre-existing knowledge in a more effective manner. However, Bransford (1984) argues that advance organizers should be written differently depending on whether they are to be used for schema activation or schema construction. He states that an advance organizer can be effective if the learner has already acquired the necessary schemas for the given problem. However, it will not be of much help for schema construction.
Bruner’s (1966) cognitive structure (i.e., schema, mental models) is consistent with schema-based learning in that it describes learning as an active process in which learners construct new ideas or concepts based on their existing knowledge. The learner does not passively respond to stimuli but actively selects and transforms information, constructs hypotheses, and makes decisions based on his or her cognitive structure. Bruner asserts that the cognitive structure supports the learner, actively generates meaning for real-world experiences, and allows the individual to process information in a meaningful way. He emphasizes that the learning situation should help learners to actively reorganize the new information, allowing them to build on existing knowledge in a meaningful way and use the newly gained meaningful knowledge effectively in the future.
Wittrock’s concept of generative learning asserts that the learner actively constructs the whole process of learning. Wittrock (1991) states with regard to the generative model that “learners must construct between stored knowledge, memories of experience, and new information for comprehension to occur.” Accordingly, an important aspect of generative teaching is knowing learners’ preconceptions, how to modify these preconceptions, and how to induce them to generate a new model of the phenomenon by revising or transforming their models (Wittrock 1991).
Cognitive load theory assumes that learning consists primarily of the acquisition of schemas. Sweller (1988) understands schemas as the cognitive structures that compose an individual’s knowledge. Learning requires a change in the schematic structures of long-term memory and is demonstrated by performance that progresses from slow and difficult (novice-like) to smooth, fast, and effortless automation (expert-like). The change in performance occurs as the learner’s schemas are increasingly associated (activated) with the learning material. Consequently, it requires instructional techniques that take the optimal level of cognitive load into account and do not interfere with schema acquisition (Sweller 1988).
In the 1980s, many knowledge-based AI systems used schemas (so-called knowledge packets) as the fundamental basis for computational models of understanding, planning, and problem solving. In this field of application, schemas usually serve as structured beliefs which involve causal-predictive cycles of action and perception. Such models have been used to interpret basic speech acts and linguistic expressions in an agent’s physical environment in terms of grounded schemas (Roy 2005). Another application of schema-based learning is generalization by means of knowledge chunking (DeJong and Mooney 1986). The idea of generalizing with schemas was closely related to the development of explanation-based systems, which are supposed to be capable of learning new schemas which can be used to generalize examples (O’Rorke 1984). This form of generalization is possible because schemas themselves are generalized knowledge chunks. However, up to now, only very few of these artificial systems are capable of generating their own schemas.