In this paper we show how simple simulation scenarios can be used to develop and test foundational functionalities of cognitive architectures, exemplified with the SiMA architecture. We present an interdisciplinary methodology that considers the challenges in capturing and evaluating basic functionalities of the human mind. In this regard, we structure and concretize assumptions from various disciplines and show how we evaluate their plausibility in a consistent model, using parametrized simulations. 1 ARS (Artificial Recognition System) was renamed to SiMA.
The computational approach to examine the human mind
provided a powerful methodology of research. When the
examination of information processing systems (such as the
human mind) is at stake, computer scientists are particularly
suitable to contribute their experience. Nevertheless,
computer scientists often still approach problems of
Cognitive Science in a classical AI way. This is especially
the case regarding interdisciplinary: instead of concretizing
models of the human mind from other disciplines into a
consistent and testable form, often own models that suffice
computational criteria (such as efficiency) are developed. In
this regard computer science stays behind its possibilities in
contributing to understand the human mind. A
counterexample is to use the computational methodology in an
approach of synthetic psychology (Braitenberg, 1986).
Similarly, computational models often focus on simulating
high-level cognition without considering their foundations,
such as motivation and emotion. We propose a more natural
approach in considering the foundations of cognition in a
unified cognitive architecture that harnesses the possibilities
given by computational simulations and is able to provide a
unified tool to test assumptions and their relationships to
each other. We will use superficially simple simulation
scenarios to guide our development and test the resulting
model. On the one hand this considers that most of humans’
behaviour is covered by every-day capabilities (what
A good overview and classification of cognitive
architectures are elaborated in
Prominent examples of symbolic architectures are SOAR, EPIC, ICARUS and NARS. Examples of emergent architectures are IPCA, Cortronics, NuPIC, and NOMAD. ACT-R, CLARION, LIDA, DUAL, Polyscheme, 4CAPS, Shruti, and Novamente can be regarded as hybrid architectures. For a description of these projects see, for instance, Duch, Oentaryo & Pasquier (2008).
ACT-R
In SOAR
Furthermore, LIDA
In ICARUS
Rated according to the scheme sketched above the SiMA architecture is a hybrid one. It defines three layers, where the lowest one comprises the neural activities, i.e. the sensor and actuator activities (see Figure 1, the leftmost block). The second layer has to build neurosymbols from the neural input and in the other direction neural actuator signals from the symbolic results of the topmost layer, the psyche, which is understood as a symbol processing machine.
sexual drives self-preservation drives external perception body perception ascnenteaunenuastrdsotao,rlrss snyz(emadutebiro-oo)nl-iaction drive track/ unconscious perception track/ unconscious
Super-Ego rules
defense mechanisms
Some specific key features define the SiMA approach. First and foremost it is a functional model, i.e. it follows a generative approach with the focus on describing functions that generate behavior instead of building a behavior model. This enables a generic and flexible model. Another feature is the layered description of human information processing. The principle here is to use appropriate means of description for different aspects of a systems, e.g. the neuronal layer should be described with other means than the psychic layer. In developing such a model we use a holistic and unified approach, which considers a consistent and coherent description of all key aspects of human information processing. The consideration of these key features is only possible by following a bionic and hence interdisciplinary approach.
The impetus for SiMA was the challenge, to design a
control unit able to cope with ambiguous situations, such as
the security monitoring of an airport or the cooperation of a
robot with human co-workers. The artificial system should
have the same “feeling” for a situation as a human would
have. The only way to gain this could be the bionic
approach. So a holistic theory of the human mental
apparatus as the control unit of the human body
The challenges in capturing the functionality of the
human mind in an interdisciplinary collaboration using
computational simulations pose special requirements on the
methodology in developing and evaluating the SiMA model.
The question here is, how to translate assumptions about
human mind’s functioning from other disciplines in a
deterministic and testable simulation model? In the SiMA
project case-driven agent-based simulation
The first step in case-driven agent-based simulation is the analysis of the model requirements. Here, we use a casuistic approach, where the behavior and underlying assumed psychic processes in a concrete case, e.g. a hungry agent perceiving a food source and another agent, are described in a narrative way. But our experience in interdisciplinary collaboration showed that a textual concretization and structuration is needed to use such exemplary case as a basis for further development of a causal and deterministic model. Overall, the procedure of case-driven agent-based simulation consists of following steps (also see Figure 2). exemplary form, e.g. regarding motivations and decision making in a concrete internal and external situation.
The exemplary case primarily serves as a platform for interdisciplinary collaboration that facilitate the discussion between researchers, which often use different approaches and vocabularies. Hence, the usage of a concrete case supports bridging the disciplines and enhances the understanding.
The exemplary case at hand (called “Adam is hungry”) describes a simplified gent-object interaction. The initial situation is given by hungry Adam, the agent with the SiMA architecture, Bodo, a passive agent, and a Viennese Schnitzel as a food source. The exemplary case describes abstractly how Adam’s motivations, represented by drives, get in conflict with perception and social norms. And how mediating psychic processes finally decide his actions. In short, Adam is confronted with choosing to eat, share, give the Schnitzel or even beat Bodo. But under which external and internal conditions does he choose the respective alternative actions? A deterministic description is needed. Generally, to use the exemplary case as a point of departure for model development, some criteria must be considered. These are especially the explication of assumptions and requirements, and the consideration of a consistent and deterministic description with a concrete focus (e.g. motivation and decision making). Therefore the exemplary case is transformed into, structured simulation case.
A focus in analyzing and transforming the exemplary case into a simulation case is an analysis of the data that determine the agent’s behavior. Following a functional approach we also focus on how a change in these data determinants would lead to a behavioral change. We distinguish four groups of determinants: the agent’s experience, personality factors (as simplifications of memories and body functionalities), the environmental state, and the agent’s initial internal state (given by drives and emotions).
The simulation case for the described exemplary case is sketched in Figure 3, with the standard scenario of eating the Schnitzel, and the alternative scenarios of beating Bodo, and sharing or giving the Schnitzel to Bodo. As sketched, the personality factor “neutralized intensity”, which indicates the strength of the defense and secondary process (see below), plays a key role in the selection of the scenario.
The transformation into a structured simulation case follows use-case-based requirements analysis in software engineering. Data determinants represent pre-conditions, the description of an agent’s final internal state and selected action represent post-conditions. For the standard scenario, eat, the inner processes that generate the post-conditions from these pre-conditions are described step-by-step. We also have to track and justify every possible behavior of the exemplary case (e.g. share, beat). That is, for the alternative scenarios we only describe how the change of data determinants would lead to an alternative behavior.
Overall, this structuration enables a fine-grained requirements analysis, the development of a causal model and its evaluation.
After developing the model (see sections below for an overview) it is tested using the simulation case as a test plan. In particular we parameterize the simulation according to the scenario’s data determinants and observe if the functions generate and data determine behavior as expected (see Chapter Calibration and Simulation). We do not only validate the behavior, but also how the behavior is generated and determined, e.g. how emotions and drives evolve and influence the agent’s decision. If the agent behaves unexpected or the data visualization indicate wrong assumptions, based on an analysis on different levels, we have to conduct another iteration of the procedure (see feedback cycles a, b, and c in Figure 2). Possibility a and b indicate that the inputs form other disciplines (psychoanalysis, neuroscience) are interpreted and transformed wrongly or implicit requirements emerge during implementation (implementing a model helps to understand and specify it). Possibility c may be caused by inconsistencies in an underlying theory or between different theories. This feedback helps to sharpen theories from other disciplines precisely.
This evaluation methodology enables us to test our model’s predictability and plausibility; in particular, the validity of the case’s assumptions and if the specified data determine the expected behavior (change).
In SiMA, the primary process represents unconscious data processing. It is characterized through fast and immediate processing of data that is close to sensor values. Its logic can be well described by the rules that apply on associations between data structures. There are two rules of the forming of associations: similarity and simultaneousness. This means that things that are similar are likely to form associations as well as objects that occur at the same time or within a short time frame.
The inputs of the primary process are defined by the homeostasis of the body, the body perception and the external perception. Homeostatic values are symbolized into drive tensions, which are a mean for intensity of bodily need. In the Drive Track in Figure 1, drives are created from the drive tension and extended with a drive object and a drive aim. The drive object is the external object, which is able to satisfy a drive and the drive aim is the action that has to be taken to satisfy it.
External perception and body perception are symbolized
and define the input of the Perception Track in Figure 1.
Here, based on perceived features, internal representation
(so-called percepts) of objects are inferred. Through the
property of simultaneousness, these objects form a
perceived image that represents the current situation.
Through the property of similarity, similar stored situations
as stored images are then activated. In the stored images,
memorized emotions are associated that reminds the system
of a certain emotional state. Together with the drives, they
generate the current emotions
Internalized rules Defense mechanisms: - reduce input data - detect and resolve conflicts in input data Primary process Secondary process
At the verge of the primary process to the secondary process are the defense mechanisms located (see Figure 5). Defense mechanisms are a kind of filter mechanism. The two tasks of the defense mechanisms are, firstly, to reduce the data which flow from the primary process to the secondary process and, secondly, to detect and resolve conflicts in input data. In order to process the first task (data reduction), the data are assessed by emotions and the focus of attention is set on specific data with a high level of activation. To process the latter task, to detect a conflict, the defense mechanisms have access to an internalized rule base, the Super-Ego-Rules. And in order to, eventually, resolve a conflict, the defense mechanisms can repress input data or alter them before the defense mechanisms pass them on to the secondary process. Which defense mechanism is chosen, depends on personality factor “conflict tension” (severeness of a conflict) and the stage of development of the personality of the software agent.
The secondary process is responsible for the preconscious/conscious processing of data. Its main task is to take a decision about an action based on the inputs from the primary process. However, different to the primary process, more extensive associations of data structures are possible. Data structures are extended with a word, making it possible to communicate the information to a received outside of the system. Also, temporal and hierarchical associations may be used, making it possible to order things. At the beginning of the secondary process, activated stored images, which were independent in the primary process are formed into sequences called acts. Acts define events and the actions necessary to be taken to get from an event to another.
Decision making in SiMA can be divided into two stages as seen in Figure 6. In each stage, similar process steps are taken: first to limit the number of options, considering all available options. Second, the options passing the first stage have access to more system resources and one of them is finally selected.
The first step in the decision making process is to extract the possible options that the system can develop and act on. It is the start of the Selection of Need track of Figure 1 and Figure 6. This is done through the creation of possible goals (“propose options” in Figure 6) from the acts or from perception. Drives from the primary process become drive wishes, which are one of the motivations to do something in the system. They define the desired external object, the preferred action and the importance to reach it. Emotions are transformed into feelings that can also be used to emphasize or to avoid certain situations.
After a general initialization with a basic effort analysis, the possible goals are evaluated regarding the possibility to fulfill a certain drive wish, an emotional state, and social rules (evaluate goals” in Figure 6). Based on the available system resources one or more possible goals are selected for further processing (“select option” in Figure 6).
The selected possible goals are the options that the system has at the moment. In the Selection of Action track of Figure 1 and Figure 6, possible action plans are generated and evaluated for each of them. Then, one option is selected and executed.
Decision making of the secondary process is a deliberative process in contrast to the primary process.That is, the options of the system can be processed during multiple cycles without any external actions. The system can reason about several options in sequence before taking a decision. Internal actions are used to perform analysis of options and to execute queries to the knowledge base that modify the internal state of the possible goal.
The modules in SiMA encapsulate functionalities of the human mind and are developed independently, following a black-box approach. Meaningful integration tests for these modules require a level of knowledge about module interaction, which is not available, due to the high number of modules, parameterization options and their functional structure. Therefore we keep integration testing to a minimum in favor of system testing, using exemplary cases. Calibration is performed in various steps on each scenario. First the environmental situation (Adam, Bodo, and the Schnitzel) is modelled as the most basic layer of calibration. Next, the drive situation is modelled and memories are created to match memorized actions to drives, according to the simulation case description. Where needed, the defense mechanisms are modelled and harmonized with the drive situation. Lastly, the acts are modelled and associated to the memorized actions. Each step could, and often did, require previous steps to recalibrate to allow modeling according to the description. This resulted in a calibration strategy similar to a waterfall model with feedback.
As mentioned, we validate our model via test scenarios in the MASON simulation framework. This chapter summarizes the results of these simulations. We compare the agent’s behavior and internal state to the expectations defined in the simulation case. The internal state of the agent is checked via data visualizations.
The simulation scenarios are designed to show the capabilities and impacts of the functional modules. Exemplary case 1 is focused on the primary process, specifically the interaction between perception, drive state and defense mechanisms. The secondary process focusses on following the memorized action sequences. In each scenario, the agent can choose between four plans: EAT, BEAT, GIVE and SHARE2. The initial environmental situation is also shared among scenarios. The blue lines indicate sight ranges, the green Agent will be referred to as Adam, the red agent as Bodo and the round shape between them as Viennese Schnitzel. 2 Plans are written in capital letters to distinct them from actions.
The first chart in Figure 8 describes the behavior of Adam during the standard scenario (eat the Schnitzel). The first column shows Adam’s current plan. The combination of high hunger drive (see Figure ), the perception of a Schnitzel and the memorized satisfaction for eating Schnitzel, make Adam initially follow the plan to EAT. After the Schnitzel was consumed, Adam switches to the BEAT plan, as it fits the new perception (Bodo; no Schnitzel) and new drive state.
Figure visualizes the changes in iAndamdestadilr.iveAsdtaamte Figure 9: Drives: hunger in green starts with high (aggressive) and red (libidinous). hunger. While he eats, the hunger drops, since eating the Schnitzel changes Adam’s body state which the drives represent. In time, the hunger subsided below the sexual drives, which started out low but steadily increase. The stamina drives (Figure 9 in blue and cyan) represent Adam’s need for relaxation and changed in response to Adams exhaustion while approaching the Schnitzel (first two peaks) and Bodo (third peak).
This scenario differs from the standard scenario in its initial drive state. Adam starts with higher, faster increasing sexual drives and low hunger. The BEAT plan is memorized with the highest satisfaction for the sexual drives and is associated with the current emotional state (see Figure 10).
Beating reduces the anger and causes a short peak of joy.
These scenarios use a defense mechanism to alter Adam’s behavior away from the current drive demands. This is achieved by the drive mechanism “sublimation”, which changes the valuations of the possible actions associated to the hunger drive, away from their memorized satisfaction values. Due to their similarities they are discussed together. The third chart in Figure 8 shows Adam’s behavior during the give scenario and the fourth chart shows Adam’s behavior during the share scenario.
In both scenarios, the drive situation is similar to the standard scenario, with the hunger drives dominating. A super-ego rule is introduced to create a conflict between high hunger drives and perceiving Bodo with the Schnitzel. The conflict is resolved by subliming the action eat, which is chosen due to its high memorized satisfaction for the hunger drive (see description of standard scenario eat), with the action give or share. In both cases Adam will follow the corresponding plan, as the sublimed action now promises the highest satisfaction for the hunger drive.
The calibration of drives makes for a good point of departure for calibrating the SiMA model. One possible problem at this stage is that the drive situation also influences the perception. Calibrating the perceived reality is normally done straight forward by directly specifying the simulated world, but extreme drive states influence the selection of memories for perception, since the agent also activates memories that are that satisfy current drives. After modelling the base behavior, the defense mechanism can be used to fine tune the agent. They can alter the influence other primary process modules have on the decision. In the current simulations, the defense was only used to model alternative behavior in situations that match certain criteria (e.g. sharing when Bodo is next to the food source). More elaborated uses of this tool are thinkable as the defense is capable of influencing the impact other primary process modules have on the decision (e.g. fine tuning the drive state via super-ego rules). This ability to influence, and even alter, the results of other modules, also makes the defense harder to calibrate, since changes in this module may affect many others and vice versa.
We showed how a narrative exemplary case that exemplifies assumptions and requirements for a cognitive architecture is structured into a deterministic description and test plan. The resulting model is evaluated in simulations, where we showed how the interplay of various decision factors, such as drives, emotions, social norms, results in an adapted decision for an agent’s internal and external state. Using a functional model we emphasized on how behavioral change is generated without changing the model.
The presented evaluation using simulation cases can be regarded as a calibration that validates the transformation of the SiMA model into a software implementation. However, such calibration is only a first step to demonstrate the model’s consistency and plausibility, with the next step being to observe if unspecified parameters would generate comprehensible results without model adaptions.
In the end such methodology is a premise to apply the model in a specific domain, e.g. as a decision support tool for marketing strategies (as currently approached), where we will additionally test our model against empirical data.