2.2.1. Long-term learning

Long-term learning affects the way existing experiences are used during memory construction over a period of time. It deals with the adjustment of an experience's state that facilitates its future utility. This form of learning is effected through the experiential grounding of an experience that has been used in the current situation to effect actions onto the environment. An experience that has been grounded frequently is more likely to be used in subsequent memory constructions. The effects of grounding are modeled by the notion of levels of processing (Fig. 1).

The process of grounding entails the embedding of an artificial agent into its environment such that its behavior and the mechanisms, representations, etc., underlying this behavior are intrinsic and meaningful to the agent itself, rather than dependent on an external designer or observer (Ziemke, 1999). The meaningfulness of these mechanisms, representations, and so forth, is determined by their usefulness in assisting the agent in attaining any of its goals (Nehaniv, 1999). Although symbolic groundings deal with the provision of meanings for symbolic representations, experiential grounding deals with the provision of meaning for the experiences processed by the artificial agent. (Hereafter “grounding” refers to experiential grounding as described in this section.) For an experience to have meaning for an agent, the consequence of actions based on this experience must have an impact on the agent itself (Nehaniv & Dautenhahn, 1998). Underlying this meaningfulness of experience is the applicability of the experience through its usage as a basis for effecting actions. The interaction between the agent and the environment plays an important role in experiential-based grounding and has an effect on the use of existing experiences processed by the agent. An agent, with its current experiences, creates a particular view of the environment. If this view matches the other view created by agent through its interactions with the environment, it becomes useful in predicting the effects of a course of actions. The experience underlying this prediction is labeled useful and is grounded. This process of grounding models the effect of long-term learning.

Different levels of processing model different groundings of experiences based on their previous applications. A highly grounded experience has a high level of processing and is used more readily in the subsequent operation of the memory system. The different levels of processing are divided into two regions according to a predefined value of grounding to demarcate short-term memory (STM) and LTM (Fig. 1). An experience starts off in STM after its first creation by memory construction. This experience progresses toward LTM as it gains groundings through its applications in subsequent memory constructions. As its usage increases, the groundings of other experiences reduce, causing them to move toward STM. Experiences in STM are purged out of the system when required. These behaviors treat STM as a gateway for experience to get into LTM or out of the memory system.

Note that the notions of grounding and levels of processing determines which experience is given higher priority for use in the current memory construction. Experiences residing in both STM and LTM are available for use during memory construction. STM operating as a gateway is only applicable to the “movement” of experiences among the different levels of processing according to the groundings they received.

2.2.2. Constructive learning

The experiences processed by an agent affect the way in which it interacts with the environment (Lueg & Pfeifer, 1997). These experiences provide a basis for determining the actions to be effected in the current situation (Rylatt & Czarnecki, 1998). A commonly accepted notion about expertise is that experts differ from novices in terms of the number and quality of experiences they have in a particular domain. Useful experiences acquired over time form the basis for evaluating and refining complex knowledge organization (Goker, 1997; Seifert et al., 1997). This has the implication that the contents (knowledge and experiences) of a memory system are dynamic in nature and vary accordingly as the agent encounters new situations.

Constructive learning models these dynamics of a memory system based on the constructivism view of learning. This view of learning adopted here is derived from two fundamental principles of constructivism (Tait, 1997).

• Learning involves an active construction of a conceptual knowledge base of the learner (Bruner, 1966; Piaget, 1971; Jonassen, 1991).
• Learning is reflective in nature. It builds upon and develops the existing knowledge and experiences of the learner (Bruner, 1966; Ausubel, 1968).

Constructive learning uses similar experiences as bases to create new experiences during memory construction. Elements of a past experience are mapped into elements of the current situation to form a coherent memory structure for performing actions onto the environment or to the agent itself. The effect of constructive learning is a change in structure of the memory system through a reconfiguration of existing experiential components or the addition of new ones. Constructive learning occurs whenever a constructed memory is added into the memory system as a new experience.

2.3.1. Representation of design experience

A design situation models a particular state of interaction between a design agent and the environment at a particular point in time. It is a snapshot of all the variables that define the internal state of the agent and the external state of that part of the environment that the agent is interacting with. The expectation and interpretation of the agent about the environment may or may not be in agreement with each other within a design situation. This agreement is based on the comparison of the behaviors contained within the expectation and interpretation of the environment. A disagreement implies that there is a difference in terms of either behavioral attributes or attribute values and this forms the basis for actions.

A design experience models the transition of state from one design situation to another situation in which the expectation is anticipated to agree with the interpretation. The transition is represented by the initial and final state together with the action that facilitates the transition. Essentially, a design experience consists of two design situations and the action that transforms the initial situation into the final situation. Within the initial situation, the expectation and interpretation are not in agreement. In the final situation, the expectation agrees with the interpretation. Symbolically, a design experience (E) is represented by a collection of constituent parts:

where A is the action on the agent or on the environment; Hⁱ is the initial focused concept; H^f is the final focused concept; M is the percept; Qⁱ is the initial expectation; Q^f is the final expectation; Sⁱ is the initial sensory data; S^f is the final sensory data; Vⁱ is the initial exogenous variable; V^f is the final exogenous variable; Q^x = {F, B, S}, with Q^x denoting the initial expectation if x = i, or the final expectation when x = f, F, B, and S denoting the function, behavior, and structure of the expectation, respectively; H^x denotes the initial focused concept if x = i or the final focused concept, if x = f; S^x denotes the initial sensory data if x = i or the final sensory data, if x = f; and V^x denotes the initial exogenous variable if x = i or the final exogenous variable, if x = f.

Different aspects of designing are captured by the representation of a design experience. In this research, design information relevant to the situated function–behavior–structure view of designing are modeled. Other aspects of designing (such as design rationale) can be included in the representation of a design experience if they are to be processed explicitly by the memory system containing the experience. The use of a computational constructive memory system is limited by the representation of design experience adopted. If the required aspects of a design domain can be represented within a design experience, the proposed memory system can be used in that domain.

Focused concept. An H^x (Gero & Fujii, 2000) dictates the focus of the design agent based on a design goal. This focus determines the required functionality, and consequently, the required behaviors of the external structure according to the design goal. The function is used to bias sensation and perception (described later) while the behavior is used when the interpretation is evaluated against the expectation of the environment. A design prototype (Gero, 1990) is used to represent a focused concept. This prototype represents the design knowledge used within a design experience. A design prototype can either be retrieved from the memory system or derived from existing prototypes. External processes such as analogy, combination, mutation, and designing from first principles (Rosenman & Gero, 1993) and reformulation processes (Gero, 1998b) that deal with reparameterization and parameter expansions can be employed for this purpose. These operations are instantiations of the general conception process (described later). Each design prototype contains procedural knowledge that defines how the prototype can be used within a design experience to achieve a specific goal. This procedure is used to derive the required action to be effected into the environment.

Action. An A is either a sequence of operations an agent performs internally on itself, or a sequence of operations performed externally on the environment (i.e., on the description of the design artifact). An internal action changes the behavior of the agent by altering variables representing its internal states. Long-term learning and constructive learning are instantiations of internal actions that alter the internal states of a design agent. An external action transforms the Vⁱ to V^f according to the current design goal.

Exogenous variable (external structure). The variable V^x defines that part of the environment an agent is dealing with according to its design goal. It represents a reference to the external structure of a design artifact (Gero & Kannengiesser, 2000, 2002) to be manipulated by the agent. A design agent may or may not operate on the entire external structure. What is focused upon is a function of the current focused concept. The set of exogenous variables encountered by the design agent is different for unfamiliar environments.

Sensory data. The variable S^x models the raw data extracted from V^x based on the bias set by perception. The variables Vⁱ, Sⁱ, V^f, and S^f model the state of the artifact and its associated data sensed by the agent, before and after the agent has performed the required action according to its design goal.

Concept as expectation of the environment. The design prototype representing the current focused concept contains information pertaining to the function, behavior, and structure of a design artifact in terms of a dependency network (Gero et al., 1992). This part of the prototype is extracted as a concept through conception to form Q^x (Gero & Kannengiesser, 2000, 2002) of the agent about the environment according to the current design goal.

Percept (interpreted structure). The function, behavior, and structure information from the focused concept is used in perception to bias the sensation process and structure, S^x. The function information dictates which part of the environment will be focused upon and sensed by the agent. Only parts with the same function as that of the focused concept are extracted from the environment. The variable S^x is configured into a complete structure to form percepts (M), a form of interpreted structure (Gero & Kannengiesser, 2002) during perception.

Concept as interpretation of the environment. The interpretation of the environment (Gero & Kannengiesser, 2000, 2002) is created from the percept. This interpretation is represented by a dependency network that contains function, behavior, and structure information about the interpreted structure. The behavior and structure part of the interpretation are either computed from or extracted directly from an external repository of design prototypes. An interpretation is not stored within the representation of a design experience. It is computed as and when required.

2.3.2. Processes connecting constituents of a design experience

The constituents of a design experiences are connected by a series of interrelated processes. Figure 2 illustrates an example of a design experience (E₁) based on the domain of Design for Assembly (DFA).

Processes connecting different constituents of a design experience.

Domain descriptions. DFA is an engineering design methodology that deals with the design of products for easy assembly to lower their production costs. Boothroyd and Dewhurst (1989) presented a series of Product DFA techniques that are concerned with the reduction of the cost of a product through the simplification of its design. It is proposed that the best way to achieve this cost reduction is first to reduce the number of individual parts that must be assembled and then to ensure that the remaining parts are easy to manufacture and assemble.

Figure 3 illustrates a series of mechanical assemblies used to illustrate the application of a DFA principle: reduce the number of parts (Boothroyd et al., 1982, 1994; Boothroyd & Dewhurst, 1989; Boothroyd, 1994). The main functionality of these assemblies is to fasten two plates together. From the perspective of the function, behavior, and structure of these assemblies, the common function of “fastening” is achieved through a series of behaviors, one of which (relevant to DFA) is “the number of parts.” The components (bolt and nut, snap-fit, and integrated snap-fit) represent different structures that exhibit the required behavior.

A sample of mechanical assemblies.

In terms of DFA, the integrated snap-fit assembly gives the best performance because it requires the least number of parts. This assumes that the integrated snap-fit can be produced at a cost lower than that of a snap-fit. If a snap-fit is cheaper to produce, snap-fit assemblies are preferred. During the redesign of an external structure, all bolt and nut assemblies within the structure are replaced by snap-fit assemblies or integrated snap-fit assemblies to reduce production cost. The design experience that models the replacement of bolt and nut assemblies by snap-fit assemblies is illustrated in Figure 2.

Sensation. Sensation is an information extraction process that converts the external structure (V₁ⁱ) into sensory data (S₉ⁱ) according to the focused concept (H₁ⁱ). Function, behavior, and structure information from the focused concept [FBS(1) refers to labels in Fig. 2] is used to bias perception. Perception based on the function information it now processes, biases the sensor to extract only structures that have the same required function as that of the focused concept (F). A repository of design prototypes (Szykman et al., 1998) provides the necessary function, behavior, and structure information [FBS(3)] for the external structure.

Perception. Perception structures the sensory data S₉ⁱ according to applicable function, behavior, and structure information of the focused concept [FBS(1)] to form an interpreted structure (percept) M₁₀ of the environment.

Conception. An interpretation of the environment is formed by conception. This process uses function information from the focused concept, behavior information extracted and modified (if necessary) from an external repository of design prototypes [FBS(2)], and structure information from the environment to create a dependency network that represents the interpretation. The behavior and structure information of the interpretation is modified to a form suitable for processing during this conception process. This form is dictated by the function, behavior, and structure information from the focused concept H₁ⁱ [FBS(1)]. The expectation of the environment (Q₁ⁱ) is created from the focused concept (H₁ⁱ) by extracting the dependency network from the design prototype representation [FBS(4)].

Matching, mapping, and comparison of behavior. The mapping process is a simplified version of the analogy elaboration process described by Qian (1994). Mapping is basically a transformation process used to identify structural correspondences between the expectation and interpretation. The basic aim is to find structure mapping (correspondence) through function, behavior, and structure matching (indexing). In general, matching implies that two (function, behavior, or structure) variables are identical at an abstract level. Mapping implies that two variables have the same associated functions or behaviors. Mapping is possible only after matching has been carried out successfully. When two functions match, the associated mapping for behavior and structure can be found. If two behaviors match, only structure mapping can be found. The interpretation and expectation are matched and subsequently mapped onto each other to obtain the correspondence between their structures before the required behavior is compared according to a design goal.

Derivation of action. When matching and mapping have been completed successfully to produce a structural correspondence between the dependency network of the expectation and interpretation, the procedural knowledge stored within the focused concept (H₁ⁱ) is used as a basis for deriving the action (A₁) on the environment.

Completion of design experience. The action A₁ is effected onto the environment to create a new structure (V₂^f). The sensation, perception, conception, and matching process is repeated to form the final situation.

Initial and final situation. In the final form, the experience is made up of an initial design situation, an A₁, and a final design situation. The bolt and nut assemblies within the initial external structure (V₁ⁱ) are replaced by snap-fit assemblies in the final external structure (V₂^f).

3.3.1. IAC formulation

The standard formulation of the IAC network (McClelland, 1988) is used as a base upon which learning mechanisms are incorporated. In an IAC network the total input for node j(t_j) is

where α represents the external input scaling, e_i represents the external input, w_ji represents the weight to node j from node i, and a_i represents the activation of node i. During cueing, if the probe contains the node j, e_j is set to 1.0. α is set to a global value of 4.0 for the entire network.

After the total input has been calculated for all the nodes, activations are updated according to the following equations (O'Reilly & Munakata, 2000):

where Δa_j denotes the amount by which to change a_j.

This formulation differs from the original IAC with the additional parameter step (s) for controlling the rate at which the node activations are changed: s is set to a global value of 0.1 for the entire network; a_max is set to 1.0, a_min to −0.2, rest (r) to −0.1, and decay (d) to 1.0. The network starts off with all nodes having an initial resting activation value of −0.1. Positive links are set to 1.0 and negative weights are set to −1.0. These settings are used for the initial network before any learning occurs. The same parameter values are also used for all newly created experiences.

During memory construction, a probe is created within the IAC network as a cue for memory construction. The contents of this probe consist of a new property node that models the external structure from the environment and an existing property node that is reused for modeling the current goal of the agent for memory construction. Once the probe has been created, its property nodes are clamped (given an external input) and the network is cycled for a predefined number of times. For each cycle, the activation value for each node is updated according to the above equations. When all the activation values have settled into an equilibrium state, the resulting instance nodes and their associated property nodes that are activated with positive values, highlight similar design experiences that can be used in the current situation. Each activated experience has an associated activation value based on its usefulness in previous design sessions. A reference to the activation value of an activated experience refers to the positive activation value of the instance node representing that experience within the extended IAC network.

3.3.2. Learning mechanism

The original IAC network (McClelland, 1981, 1988; Rumelhart & McClelland, 1986) does not provide any learning mechanism. Learning can occur through:

• constructive learning via changing the structure of the network; or
• long-term learning via grounding through the adjustments of weights between existing nodes of a design experience.

Constructive learning via system restructuring. Computer implementations of constructive learning are exemplified in the modification of neural network structures (Mandic & Chambers, 2001). Constructive learning in neural networks deals with the changing of architecture or interconnections within the network. There are two basic classes of constructive learning: network growing and network pruning. In the network growing approach (Hoehfeld & Fahlman, 1992), the network starts small and grows as it learns. Network pruning starts with a large network and reduces its size until the desired ratio between accuracy and network size is reached (Reed, 1993). The idea behind these learning mechanisms is similar to the concepts of ontogenic networks (Fiesler & Beale, 1997) where the topology of the network is changed during learning.

Constructive learning within the IAC via network growth can occur with the addition of new instance nodes that model new experiences (constructed memory) created via memory construction (Fig. 6). New property nodes are created and added into the system according to new external structures from different environments (Interpretation I, Reinterpretation I, and Reinterpretation II) and according to the reconfiguration of existing property nodes (Reinterpretation I). Constructive learning via network pruning occurs through the forgetting process based on the levels of processing of different experiences. In this processes, instance nodes and property nodes that are not shared among other instance nodes are removed from the IAC network.

The IAC network before and after various forms of constructive learning.

Long-term learning via weight modification. Computer implementations of long-term learning are exemplified in the modification of weights in neural networks (Haykin, 1998; Medler, 1998). Long-term learning in the IAC networks deals with the changing of interconnection weights within the network of nodes and links through a grounding process. Grounding causes the positive links of a grounded experience to increase in value while reducing the positive link values of all other experiences through a decay process. Only the weights of excitatory (positive) links are allowed to change. Negative links are not changed locally as their global settings are used to control blending.

The learning mechanism is formulated as (Medler, 1998)

where Δw_ij denotes the amount by which to change the link between node i and node j, β denotes the learning rate, δ denotes the weight decay factor, and w_max denotes the maximum weight value. This formulation of learning is similar to that for computing activation values of nodes. The learning process occurs as a separate process only after memory construction has taken place.

The learning processes based on weight adjustments in the IAC network is carried out by clamping the property nodes of the experience to be grounded and cycling the network according the formulation for learning above. This emphasis of a particular experience strengthens the interconnections between the nodes involved. Whereas it is possible to overtrain other neural networks, it is not possible to overtrain the extended IAC network via grounding of experiences based on their applications. The activated experiences are always evaluated according to the current situation through sensation, perception, conception, matching and mapping, and comparison of behavior.

3.3.3. Implementation of STM and LTM

STM and LTM are implemented as a multilayered IAC network (Fig. 7). The separation of experiences into different layers is based on their levels of processing. The common interconnection weight between the instance node and its corresponding property nodes of an experience provides the basis for these levels. Experiences with high interconnection weights are placed close to the left of Figure 7 to indicate that they are more durable within the memory system. Experiences with low interconnection weights are placed toward the right of Figure 7. A threshold value for interconnection weight is set up to delineate LTM and STM. This threshold can be visualized as a cutting plane that separates LTM from STM among the different layers of the network.

The demarcation of STM and LTM. The number of levels for STM and LTM is not limited to those shown.

New experiences start off in STM (right side, Fig. 7). As the grounded usage of an experience increases, it is repeatedly presented to the memory system to strengthen its interconnection weights and moves toward LTM (left side, Fig. 7). All experiences in STM and LTM are subject to a decay process. If the subsequent memory construction process does not reinforce an experience, its interconnection weights will decay and its will be moved into the STM region. When experiences are removed from the system by a forgetting process to increase computational performance, the experiences within STM are purged from the system first. The separation of experiences into STM and LTM is only critical during this removal process.

4.2.1. Different operating modes in the same situation

The purpose of Experiments 1 and 2 is to highlight the effects of grounding different experiences on the operating modes of separate memory systems (residing in different agents) in a common situation. Two experiments that produce the same response are performed to illustrate the internal changes within different memory systems before the threshold that produces a different response is crossed. The two experiments, each with two runs of memory constructions, are

where P₁ has a cue based on {V₁₀₁ⁱ, H₁ⁱ, and H₁^f}, P₂ has a cue based on {V₁₀₂ⁱ, H₁ⁱ, and H₁^f}, and P₃ has a cue based on {V₁₀₄ⁱ, H₁ⁱ, and H₁^f}. The first run in each experiment represents a common memory system used in both experiments. The effects of grounding E₈₀₁ and E₁ on the operating modes of separate memory systems in the same situation are illustrated. Among these memory systems,

• the same response within the same situation is indicated by the same experience being activated with the maximum activation value when their respective agents encounter the same situation;
• different operating modes are indicated by different maximum activation values (in terms of their magnitudes) of the common response (experience) from each system when their respective agents encounter the same situation; and
• the same situation encountered by their respective agents are indicated by the use of the same cue.

Experiment 1. Figures 9 and 10 illustrate the internal states of two memory systems based on the Run 1 (P₁) and 2 (P₂ → G_E801 → P₁) of Experiment 1. The response of each system within a common situation (modeled by P₁ at the end of each run) is the same (E₈₀₁ having the maximum activation value) but each memory system operates differently to arrive at this response. This is indicated by the different activation values of E₈₀₁ (0.845 at the end of Run 1, Memory State 1, Fig. 9; and 0.849 at the end of Run 2, Memory State 4, Fig. 10) within the two memory systems. The difference in value is significant. The increase in activation is due to the additional grounding of experience E₈₀₁ thus increasing its likelihood of usage within the corresponding memory system.

Different states of the memory system for Experiment 1, Run 2.

Discussion of results: Experiment 1. The two runs of the experiment illustrate the memory construction histories of two separate agents through a series of memory states. Although both agents start off with the same set of experiences (Fig. 8), the second agent (with history modeled by Experiment 1, Run 2) learned about the usefulness of experience E₈₀₁. This exposure to the use of E₈₀₁ causes its memory system to respond with E₈₀₁ at a higher activation value when both agents are placed in the same situation later.

When interpreted within the domain of DFA, the behaviors of the two agents can be interpreted as follows: both agents start off with the same set of experiences that deals with redesigning according to DFA. Before being put into the same situation, the second agent encounters an external structure (V₁₀₂ⁱ) that it tries to redesign according to a DFA guideline that proposes the use of a snap-fit assembly (H₁ⁱ and H₁^f). This agent creates a partial design experience E₁₀₂′ according to these parameters {V₁₀₂ⁱ, H₁ⁱ, and H₁^f}, and uses it as a cue for memory construction (Memory State 1, Fig. 10). The memory system responds with an experience E₈₀₁ that proposes the use of an integrated snap-fit assembly (H₈₀₀ⁱ and H₈₀₀^f). This preference for integrated snap-fit is caused by a larger number of experiences that uses integrated snap-fit assemblies for better performance in Figure 8. The initial guideline modeled by H₁ⁱ and H₁^f is also modeled by H₈₀₀ⁱ and H₈₀₀^f, but the latter case is given preference (constructive learning allows a new focused concept to model a DFA guideline in place of another focused concept used previously, however, this part of the research is not within the scope of this paper). Through the selection of E₈₀₁, the agent constructs the required memory and restructures itself and the environment for the application of the integrated snap-fit assembly.

The selected experience (E₈₀₁) is grounded (Memory States 2 and 3, Fig. 10). This makes the likelihood of responding with the integrated snap-fit assembly even more likely in subsequent memory constructions. This higher probability of usage is demonstrated when both agents are exposed to the same situation later. Both memory systems are cued with the same partial design experience E₁₀₁′ (Memory State 1, Fig. 9, and Memory State 4, Fig. 10), and each responded with E₈₀₁ (the use of an integrated snap-fit assembly) through different emphasis to reflect its previous usage within each system.

The difference in operating modes caused by grounding gives different agents the potential to behave differently within the same situation according to what they have learned to be useful. The increase in activation value of a grounded experience in subsequent operations of its corresponding memory system is indicative of the possibility that the response of the system can be altered if sufficient grounding is provided. In terms of DFA, this means that it is possible to change the proposed use of an integrated snap-fit assembly to that of a snap-fit assembly (as the response of the system) if an experience containing the snap-fit assembly (E₁) is given sufficient grounding. This hypothesis is based on the increase in activation value of E₈₀₁ due to its grounding. If E₁ is given enough grounding, it is postulated that its activation value will be elevated to such an extent that the system responds with E₁ instead of E₈₀₁. The next experiment illustrates the boundary condition just before this happens.

Experiment 2. Figures 11 and 12 illustrate the internal states of two memory systems based on runs 1 (P₁) and 2 (P₃ → G_E1 → P₁) of Experiment 2. This time experience E₁ is grounded in the second memory system (with history modeled by Experiment 2, Run 2). The response of each system within a common situation (modeled by P₁ at the end of each run) is the same (E₈₀₁ having the same maximum activation value) but each memory system operates differently to arrive at this response. This is indicated by the different activation values of E₈₀₁ (0.845 at the end of Run 1, Memory State 1, Fig. 11, and 0.843 at the end of Run 2, Memory State 4, Fig. 12) within the two memory systems. The decrease in activation is due to the grounding of experience E₁, which causes its likelihood of usage to increase within the corresponding memory system.

Different states of the memory system for Experiment 2, Run 1 and Experiment 4, Run 1.

Different states of the memory system for Experiment 2, Run 2.

Discussion of results: Experiment 2. Although both agents start off with the same set of experiences, the second agent learned about the usefulness of experience E₁. This exposure to the use of E₁ causes its memory system to increase the activation value of E₁ when both agents are placed in the same situation later.

In terms of DFA, both agents start off with the same set of experiences. Before being put into the same situation, the second agent encounters an external structure (V₁₀₄ⁱ) that it tries to redesign according to a DFA guideline that proposes the use of a snap-fit assembly (H₁ⁱ and H₁^f). This agent creates a partial design experience E₁₀₄′ according to these parameters {V₁₀₄ⁱ, H₁ⁱ, and H₁^f}, and uses it as a cue for memory construction (Memory State 1, Fig. 12). The memory system responds with an experience E₈₀₁ that proposes the use of an integrated snap-fit assembly (H₈₀₀ⁱ and H₈₀₀^f). The proposed integrated snap-fit assembly is not used this time. The system is forced to use the snap-fit assembly based on E₁ in the current situation. This may be due to a change of conditions that renders the production cost of an integrated snap-fit assembly to be higher and thus a snap-fit assembly is preferred. Through the selection of this experience, the agent constructs the required memory and restructures itself and the environment for the application of the snap-fit assembly.

The selected E₁ is grounded (Memory States 2 and 3, Fig. 12). This makes the likelihood of responding with the snap-fit assembly even more likely in subsequent memory constructions. This higher probability of usage is demonstrated when both agents are exposed to the same situation later. Both memory systems are cued with the same partial design experience E₁₀₁′ (Memory State 1, Fig. 11, and Memory State 4, Fig. 12). Although the second memory system responds with the same suggestion of using the integrated snap-fit assembly, this time it is given less emphasis, as indicated by the decrease of activation of E₈₀₁ (from 0.845 in Run 1, Memory State 1, Fig. 11 to 0.843 in Run 2, Memory State 4, Fig. 12). The activation of E₁ has increased (from 0.839 in Run 1 to 0.843 in Run 2) due to its previous grounding. When the second memory system of both Experiments 1 and 2 are compared, the final activation of E₈₀₁ is also observed to decrease (from 0.849 in Experiment 1, Run 2, Memory State 4, Fig. 10, to 0.843 in Experiment 2, Run 2, Memory State 4, Fig. 12). This is indicative of the change within the two memory systems due to the learning of different experiences.

The behavior of the second memory system in this experiment is starting to confirm the hypothesis stated in Experiment 1 that it is possible to alter the response of the system if sufficient grounding is provided. The activation value of any activated E₁ can be increased through subsequent grounding to enable it to gain maximum activation and to alter the response of the system in the same situation subsequently. This alteration of response is demonstrated in Experiment 4.

4.2.2. Different operating modes across different situations

The purpose of Experiment 3 is to highlight the effects of grounding on the operating modes of separate memory systems (residing in different agents) in a common situation through the grounding of different experiences across a different situation. The emphasis is on showing the different operating modes of separate memory systems caused by the learning of a different experience in another situation. The experiment is performed without crossing the threshold of response change, to produce similar responses to illustrate the internal changes within each memory system. One experiment with two runs of memory construction is considered.

where P₄ and P₆ have cues based on {V₁₀₅ⁱ, H₁ⁱ, and H₁^f} and {V₁₀₇ⁱ, H₁ⁱ, and H₁^f}, respectively; P₅ has a cue based on {V₁₀₆ⁱ, H₁ⁱ, and H₁^f}; and P₇ has a cue based on {V₁₀₉ⁱ, H₉₀₁ⁱ, and H₉₀₁^f}. The effects of grounding E₉₀₁ in a totally unrelated situation on the operating mode of a memory system in the same situation are illustrated.

Experiment 3. Figures 13, 14, and 15 illustrate the internal states of two memory systems based on Run 1 (P₄ → G_E801 → P₅ → G_E801 → P₆), and Run 2 (P₄ → G_E801 → P₇ → G_E901 → P₆) of Experiment 3. The response of each system within a common situation (modeled by P₆ at the end of each run) is the same (E₈₀₁ having the maximum activation value) but each memory system operates differently to arrive at this response. This is indicated by the reduced activation value of E₈₀₁ (from 0.853 at the end of the first run, Memory State 7, Fig. 13, to 0.846 at the end of the second run, Memory State 7, Fig. 15) within the second memory system (modeled by the second run) due to the grounding of E₉₀₁ in another situation. As E₉₀₁ is grounded, all other experiences within the system are subjected to a decay process. The effect of grounding that occurs previously (G_E801) is reduced. The increase of activation value due to grounding within the memory system is not permanent. When an experience (E₈₀₁) is not used in subsequent memory constructions, the effect of its previous grounding will eventually wear off. The memory system's mode of operation changes to reflect the effect of “use it or lose it” as a result of grounding another experience in a different situation.

Different states of the memory system for Experiment 3, Run 1. Memory State 4 is repeated in the bottom row to show the continuation from the top.

Memory States 0–5 of the system for Experiment 3, Run 2. Memory State 3 is repeated in the bottom row to show the continuation from the top.

Memory States 5–7 of the system for Experiment 3, Run 2. Memory State 5 is repeated in the top row to show the continuation from the previous diagram (Fig. 14), and Memory State 6 is repeated in the bottom row to show the continuation from the top row of the current figure.

Discussion of results: Experiment 3. This experiment shows that what an agent has found to be useful is subjected to its continual usage. If the usefulness of an experience (E₈₀₁) is not reinforced in subsequent situations, its elevated emphasis due to a previous utility will eventually wear off.

In terms of DFA, this means that the increase in emphasis on the use of a snap-fit assembly based on grounding alone (Experiment 2) is not permanent. Only through the increase in the number of new experiences created with more use of snap-fit assemblies in subsequent situations will the emphasis on the use of the snap-fit assembly be permanent. However, the effect of these new experiences cannot be seen immediately until a sufficient number of them have been created. This effect due to the number of experiences can be seen in Experiments 1 and 2 where a memory system responds with an integrated snap-fit assembly when it is cued with a snap-fit assembly initially.

4.3.1. Different responses in the same situation

The purpose of the Experiments 4 and 5 is to demonstrate that with enough grounding, a memory system can be forced to respond with another activated experience when the threshold for response change has been reached. This behavior is important to situate the operations of the memory system according to the current interactions with the environment. Two experiments, each with two runs of memory constructions are considered:

where P₁ has cue based on {V₁₀₁ⁱ, H₁ⁱ, and H₁^f}, P₉ has a cue based on {V₁₁₂ⁱ, H₁ⁱ, and H₁^f}, P₁₀ has a cue based on {V₁₁₃ⁱ, H₁ⁱ, and H₁^f}, P₁₁ has cues based on {V₁₁₄ⁱ, H₁ⁱ, and H₁^f}, P₁₂ has cues based on {V₁₁₅ⁱ, H₈₀₀ⁱ, and H₈₀₀^f}, P₁₃ has a cue based on {V₁₁₇ⁱ, H₁ⁱ, and H₁^f}, and P₁₄ has a cue based on {V₁₁₈ⁱ, H₁ⁱ, and H₁^f}.

The effects of grounding E₁ repeatedly on the responses of separate memory systems in the same situation are illustrated. The suppression of the original response E₈₀₁ by E₁ is also demonstrated. Among these memory systems,

• different responses within the same situation is indicated by different experiences being activated with the maximum activation value when their respective agents encounter the same situation;
• different operating modes are indicated by different maximum activation values (in terms of their magnitudes) of a common activated experience from each system when their respective agents encounter the same situation; and
• the same situation encountered by their respective agents are indicated by the use of the same cue.

Experiment 4. Figure 11 (reusing the results from Experiment 2, Run 1), and Figure 16 illustrate the internal states of two memory systems based on Runs 1 (P₁) and 2 (P₉ → G_E1 → P₁₀ → G_E1 → P₁) of Experiment 4. The response of the second system (modeled by Experiment 4, Run 2) is different from that of the first system (modeled by Experiment 4, Run 1) in the same situation (modeled by P₁ at the end of each run). Due to its repeated used of E₁, the same memory system used in the second agent has changed its response from E₈₀₁ to E₁. Its activation value of E₈₀₁ has decreased from 0.845 (Memory State 1, Fig. 16) to 0.843 (Memory State 4, Fig. 16) and finally to 0.841 (Memory State 7, Fig. 16), while the activation value of E₁ has increased from 0.839 to 0.843 and finally to 0.846.

Different states of the memory system for Experiment 4, Run 2. Memory States 3 and 5 are repeated in the bottom row to show the continuation from the top row.

Discussion of results: Experiment 4. This experiment confirms the hypothesis that it is possible to alter the response of a memory system if sufficient grounding is provides as postulated in Experiment 1. Although both agents start off with the same set of experiences, the second agent learned about the usefulness of E₁ repeatedly. This exposure to the use of E₁ causes its memory system to change its response when both agent are placed in the same situation. In terms of DFA, as the snap-fit assembly is repeatedly used, the groundings of the E₁ has effectively altered the suggested use of an integrated snap-fit assembly (H₈₀₀ⁱ and H₈₀₀^f in E₈₀₁) to a snap-fit assembly (H₈₀₀ⁱ and H₈₀₀^f in E₁).

This change of response within the memory system implies that even though an agent may have learned of the usefulness of an integrated snap-fit assembly in previous situations, the use of this learned usage is not absolute, as it is subjected to the current interactions between the agent and the environment. The application of knowledge (focused concepts) is situated. Situations may exist that render the use of the integrated snap-fit assembly infeasible, and these situations will cause the system to respond differently in other similar situations. The system updates what it had learned previously according the most recent interactions with the environment.

Experiment 5. Figures 17, 18, and 19 illustrate the internal states of two memory systems based on Run 1 (P₁₁ → G_E1 → P₁₂) and Run 2 (P₁₁ → G_E1 → P₁₃ → G_E1 → P₁₄ → G_E1 → P₁₂) of Experiment 5. The situation at the final memory construction stage is deliberately chosen to highlight the influence of grounding in settings that would normally cause a memory system to behave differently.

Different states of the memory system for Experiment 5, Run 1.

Memory States 0–6 of the system for Experiment 5, Run 2. Memory State 3 is repeated in the bottom row to show the continuation from the top row of the current figure.

Memory States 6–10 of the system for Experiment 5, Run 2. Memory State 6 is repeated in the top row to show the continuation from the previous diagram (Fig. 18), and Memory State 8 is repeated in the bottom row to show the continuation from the top row of the current figure.

After the first memory system (modeled by Experiment 5, Run 1) has learned of the usefulness of E₁, E₁ is grounded instead of E₈₀₁ (the activated experience with the maximum activation). This grounding is attempting to change the subsequent response of the memory system from E₈₀₁ to E₁. However, the effect of this grounding is not strong enough. When the memory system is cued with H₈₀₀ⁱ and H₈₀₀^f (contained within E₈₀₁) in the final situation (P₁₂), it responds with E₈₀₁ with a value of 0.847 over E₁ with a value of 0.840 (Memory State 1, Fig. 17).

The second memory system (modeled by Experiment 5, Run 2) undergoes three consecutive groundings of experience E₁. This grounding changes the subsequent response of the memory system from E₈₀₁ to E₁ (containing H₁ⁱ and H₁^f). The system responds with E₁ with a value of 0.847 over E₈₀₁ with a value of 0.843 when exposed to the same final situation as that of the first agent (Memory State 10, Fig. 19).

Discussion of results: Experiment 5. From this experiment, it can be seen that the response change of a memory system due to grounding is observable even when the memory system is subsequently cued with a focused concepts H₈₀₀ⁱ and H₈₀₀^f. These concepts (as cues) would normally give greater emphasis to E₈₀₁ because they are part of the experience. However, the influence of grounding E₁ repeatedly is so strong that a memory system can be made to respond with E₁ (Experiment 5, Run 2, Memory State 10, Fig. 19) instead of E₈₀₁ as it would have normally (Experiment 5, Run 1, Memory State 4, Fig. 17).

In terms of DFA, the repeated use of snap-fit assemblies (as proposed through E₁) by a agent, via its interactions with the environment, has the effect of forcing the use of a snap-fit assembly over the use of an integrated snap-fit assembly as proposed initially by the system. This effect is also evident when its memory system is cued according to an integrated snap-fit assembly. This behavior highlights the strong influence of repeated groundings on the response of the system. Without grounding, the cue based on the integrated snap-fit assembly would normally produce a response that uses an integrated snap-fit assembly, E₈₀₁ (Memory State 4, Fig. 17). With the repeated grounding of E₁, the response from the system is altered to one that proposes the use of the snap-fit assembly.