Interdisciplinary Development and Evaluation of Cognitive Architectures Exemplified with the SiMA Approach Samer Schaat, Alexander Wendt, Stefan Kollmann, Friedrich Gelbard, Matthias Jakubec (schaat, wendt, kollmann, gelbard, jakubec@ict.tuwien.ac.at) Institute for Computer Technology, Vienna University of Technology 1040 Vienna, Austria Abstract behaviour is more suitable to analyse the basic functions of the human mind. 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 State of the Art architecture. We present an interdisciplinary methodology A good overview and classification of cognitive that considers the challenges in capturing and evaluating basic architectures are elaborated in (Duch, Oentaryo & Pasquier, functionalities of the human mind. In this regard, we structure and concretize assumptions from various disciplines and show 2008; Langley, Laird & Rogers, 2009; Vernon, von Hofsten how we evaluate their plausibility in a consistent model, using & Fadiga, 2010). There, cognitive architectures are parametrized simulations. classified into three categories: symbolic, emergent, and hybrid architectures. Symbolic architectures process high- Keywords: Artificial Intelligence; Cognitive architectures; Intelligent agents; Computer simulation level symbols like objects or concepts and derive action plans thereof. In emergent architectures no symbols are Introduction processed but low-level activation signals in a network, for example an artificial neural network, are propagated. The computational approach to examine the human mind Actions emerge out of holistic structures. Emergent provided a powerful methodology of research. When the architectures are self-organizing and bottom-up structured. examination of information processing systems (such as the Hybrid architectures combine characteristics of both, human mind) is at stake, computer scientists are particularly symbolic and emergent architectures. suitable to contribute their experience. Nevertheless, Prominent examples of symbolic architectures are SOAR, computer scientists often still approach problems of EPIC, ICARUS and NARS. Examples of emergent Cognitive Science in a classical AI way. This is especially architectures are IPCA, Cortronics, NuPIC, and NOMAD. the case regarding interdisciplinary: instead of concretizing ACT-R, CLARION, LIDA, DUAL, Polyscheme, 4CAPS, models of the human mind from other disciplines into a Shruti, and Novamente can be regarded as hybrid consistent and testable form, often own models that suffice architectures. For a description of these projects see, for computational criteria (such as efficiency) are developed. In instance, Duch, Oentaryo & Pasquier (2008). this regard computer science stays behind its possibilities in ACT-R (Anderson, Bothell, Byrne, Douglass, Lebiere & contributing to understand the human mind. A counter- Qin, 2004), as a member of hybrid cognitive architectures, example is to use the computational methodology in an processes its data with the help of different modules, for approach of synthetic psychology (Braitenberg, 1986). example, a module for visual data, a module for motor data Similarly, computational models often focus on simulating (actions), a module for goals. In each processing cycle, high-level cognition without considering their foundations, production rules are matched against facts in short-term such as motivation and emotion. We propose a more natural memory. The production rule which produces outcome, approach in considering the foundations of cognition in a which is closest to ACT-R’s goals wins. unified cognitive architecture that harnesses the possibilities In SOAR (Laird, Congdon, Coulter, Derbinsky & Xu, given by computational simulations and is able to provide a 2011), a member of symbolic cognitive architectures, the unified tool to test assumptions and their relationships to processing cycle selects operators which fit the current each other. We will use superficially simple simulation problem and lead to a state which is closer to a desired goal. scenarios to guide our development and test the resulting Furthermore, LIDA (Faghihi & Franklin, 2012) is a model. On the one hand this considers that most of humans’ member of hybrid cognitive architectures. In LIDA the behaviour is covered by every-day capabilities (what Bargh cognitive cycle activates modules to filter input data, to & Chartrand (1999) called the unbearable automaticity of select actions, and to process the actions. Additionally, being). On the other hand our experience with the cognitive LIDA has several built in learning mechanisms. architecture SiMA1 (Simulation of the Mental Apparatus & In ICARUS (Langley, Choi & Trivedi, 2011), facts about Applications) (Schaat, Wendt, Jakubec, et al., 2014; the environment and objects are called percepts and beliefs, Dietrich, et. al., 2014) showed that – especially when the and rules are called skills. Skills are applied to percepts and foundations of the human mind are at stake – every-day beliefs in order to reach ICARUS’s goals. 1 ARS (Artificial Recognition System) was renamed to SiMA. 515 SiMA Approach Ego, the demands from being a social creature, and the ego, which has to mediate between the other two. This abstract Rated according to the scheme sketched above the SiMA theory is concertized in the SiMA project as a basis, which architecture is a hybrid one. It defines three layers, where is extended by contemporary theories from various the lowest one comprises the neural activities, i.e. the sensor disciplines, such as Damasio’s (2003) theory of emotions. 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 Case-driven Agent-based Simulation the symbolic results of the topmost layer, the psyche, which The challenges in capturing the functionality of the is understood as a symbol processing machine. human mind in an interdisciplinary collaboration using computational simulations pose special requirements on the sexual drives drive track/ unconscious methodology in developing and evaluating the SiMA model. self-preservation drives Super-Ego The question here is, how to translate assumptions about rules external perception perception track/ defense human mind’s functioning from other disciplines in a unconscious mechanisms body perception deterministic and testable simulation model? In the SiMA sensors and neuro- project case-driven agent-based simulation (Schaat & neural (de-) nets, symboli- Dietrich, 2014; Bruckner, Gelbard, Schaat et al., 2013) is zation actuators developed. This methodology guides interdisciplinary selection of action/ selection of need/ action conscious conscious collaboration in finding the required functions and data for a simulation model of the human mind. A combination and Figure 1: Overview of the SiMA model. adaption of casuistry, agent-based simulation and use-case driven requirements engineering proved suitable to cope Some specific key features define the SiMA approach. with the challenges of interdisciplinary collaboration and First and foremost it is a functional model, i.e. it follows a the evaluation of models of the human mind. Amongst generative approach with the focus on describing functions others these challenges are the restricted accessibility of the that generate behavior instead of building a behavior model. human mind, interdisciplinary knowledge translation, the This enables a generic and flexible model. Another feature complexity in explaining and evaluating models of the is the layered description of human information processing. human mind. The principle here is to use appropriate means of description The first step in case-driven agent-based simulation is the for different aspects of a systems, e.g. the neuronal layer analysis of the model requirements. Here, we use a casuistic should be described with other means than the psychic approach, where the behavior and underlying assumed layer. In developing such a model we use a holistic and psychic processes in a concrete case, e.g. a hungry agent unified approach, which considers a consistent and coherent perceiving a food source and another agent, are described in description of all key aspects of human information a narrative way. But our experience in interdisciplinary processing. The consideration of these key features is only collaboration showed that a textual concretization and possible by following a bionic and hence interdisciplinary structuration is needed to use such exemplary case as a basis approach. for further development of a causal and deterministic model. The impetus for SiMA was the challenge, to design a Overall, the procedure of case-driven agent-based control unit able to cope with ambiguous situations, such as simulation consists of following steps (also see Figure 2). 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 (Solms, 2009) was needed to work as the blueprint for the SiMA model. The basis was found in Freud’s metapsychology. Freud came up with two major structuring concepts for the psyche, the first and the second topographical model. In the first model the distinction is made between the primary process, where data are handled totally unconsciously according to the pleasure principle, and the secondary process with preconscious and conscious, also rational, data treatment, where additionally the reality principle gets Figure 2: Case-driven agent-based simulation. observed. From the point of view of computer science it clearly is a data model, while Freud’s second topographical Exemplary Case model is a function model. It distinguishes between the The exemplary case is a narrative description of a functions of the Id, the treatment of bodily needs, the Super- concrete case that demonstrates assumptions in an 516 exemplary form, e.g. regarding motivations and decision Overall, this structuration enables a fine-grained making in a concrete internal and external situation. requirements analysis, the development of a causal model The exemplary case primarily serves as a platform for and its evaluation. 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 Figure 3: Simulation case “Adam is hungry”. and internal conditions does he choose the respective alternative actions? A deterministic description is needed. Evaluation Generally, to use the exemplary case as a point of departure After developing the model (see sections below for an for model development, some criteria must be considered. overview) it is tested using the simulation case as a test These are especially the explication of assumptions and plan. In particular we parameterize the simulation according requirements, and the consideration of a consistent and to the scenario’s data determinants and observe if the deterministic description with a concrete focus (e.g. functions generate and data determine behavior as expected motivation and decision making). Therefore the exemplary (see Chapter Calibration and Simulation). We do not only case is transformed into, structured simulation case. validate the behavior, but also how the behavior is generated and determined, e.g. how emotions and drives evolve and Simulation Case influence the agent’s decision. If the agent behaves A focus in analyzing and transforming the exemplary case unexpected or the data visualization indicate wrong into a simulation case is an analysis of the data that assumptions, based on an analysis on different levels, we determine the agent’s behavior. Following a functional have to conduct another iteration of the procedure (see approach we also focus on how a change in these data feedback cycles a, b, and c in Figure 2). Possibility a and b determinants would lead to a behavioral change. We indicate that the inputs form other disciplines distinguish four groups of determinants: the agent’s (psychoanalysis, neuroscience) are interpreted and experience, personality factors (as simplifications of transformed wrongly or implicit requirements emerge memories and body functionalities), the environmental state, during implementation (implementing a model helps to and the agent’s initial internal state (given by drives and understand and specify it). Possibility c may be caused by emotions). inconsistencies in an underlying theory or between different The simulation case for the described exemplary case is theories. This feedback helps to sharpen theories from other sketched in Figure 3, with the standard scenario of eating disciplines precisely. the Schnitzel, and the alternative scenarios of beating Bodo, This evaluation methodology enables us to test our and sharing or giving the Schnitzel to Bodo. As sketched, model’s predictability and plausibility; in particular, the the personality factor “neutralized intensity”, which validity of the case’s assumptions and if the specified data indicates the strength of the defense and secondary process determine the expected behavior (change). (see below), plays a key role in the selection of the scenario. The transformation into a structured simulation case Primary Process follows use-case-based requirements analysis in software In SiMA, the primary process represents unconscious data engineering. Data determinants represent pre-conditions, the processing. It is characterized through fast and immediate description of an agent’s final internal state and selected processing of data that is close to sensor values. Its logic can action represent post-conditions. For the standard scenario, be well described by the rules that apply on associations eat, the inner processes that generate the post-conditions between data structures. There are two rules of the forming from these pre-conditions are described step-by-step. We of associations: similarity and simultaneousness. This also have to track and justify every possible behavior of the means that things that are similar are likely to form exemplary case (e.g. share, beat). That is, for the alternative associations as well as objects that occur at the same time or scenarios we only describe how the change of data within a short time frame. determinants would lead to an alternative behavior. 517 The inputs of the primary process are defined by the (severeness of a conflict) and the stage of development of homeostasis of the body, the body perception and the the personality of the software agent. external perception. Homeostatic values are symbolized into drive tensions, which are a mean for intensity of bodily Secondary Process need. In the Drive Track in Figure 1, drives are created from The secondary process is responsible for the the drive tension and extended with a drive object and a preconscious/conscious processing of data. Its main task is drive aim. The drive object is the external object, which is to take a decision about an action based on the inputs from able to satisfy a drive and the drive aim is the action that has the primary process. However, different to the primary to be taken to satisfy it. 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. Figure 4: Drive representation. External perception and body perception are symbolized selection of need/ conscious selection of action/ conscious and define the input of the Perception Track in Figure 1. Here, based on perceived features, internal representation First decision process Second decision process (so-called percepts) of objects are inferred. Through the Propose options Evaluate options Select option Propose options Evaluate options Select option property of simultaneousness, these objects form a perceived image that represents the current situation. Figure 6: Decision making in the secondary process. Through the property of similarity, similar stored situations as stored images are then activated. In the stored images, Decision making in SiMA can be divided into two stages memorized emotions are associated that reminds the system as seen in Figure 6. In each stage, similar process steps are of a certain emotional state. Together with the drives, they taken: first to limit the number of options, considering all generate the current emotions (Schaat, 2013) of the system available options. Second, the options passing the first stage that will be used later on in decision making. 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 Internalized Internalized rules rules Figure 6. This is done through the creation of possible goals (“propose options” in Figure 6) from the acts or from Defense mechanisms: perception. Drives from the primary process become drive - reduce input data Primary process - detect and resolve Secondary process wishes, which are one of the motivations to do something in conflicts in input data the system. They define the desired external object, the Figure 5: Overview of defense mechanisms’ function. preferred action and the importance to reach it. Emotions are transformed into feelings that can also be used to At the verge of the primary process to the secondary emphasize or to avoid certain situations. process are the defense mechanisms located (see Figure 5). After a general initialization with a basic effort analysis, Defense mechanisms are a kind of filter mechanism. The the possible goals are evaluated regarding the possibility to two tasks of the defense mechanisms are, firstly, to reduce fulfill a certain drive wish, an emotional state, and social the data which flow from the primary process to the rules (evaluate goals” in Figure 6). Based on the available secondary process and, secondly, to detect and resolve system resources one or more possible goals are selected for conflicts in input data. In order to process the first task (data further processing (“select option” in Figure 6). reduction), the data are assessed by emotions and the focus The selected possible goals are the options that the system of attention is set on specific data with a high level of has at the moment. In the Selection of Action track of Figure activation. To process the latter task, to detect a conflict, the 1 and Figure 6, possible action plans are generated and defense mechanisms have access to an internalized rule evaluated for each of them. Then, one option is selected and base, the Super-Ego-Rules. And in order to, eventually, executed. resolve a conflict, the defense mechanisms can repress input Decision making of the secondary process is a data or alter them before the defense mechanisms pass them deliberative process in contrast to the primary process.That on to the secondary process. Which defense mechanism is is, the options of the system can be processed during chosen, depends on personality factor “conflict tension” multiple cycles without any external actions. The system 518 can reason about several options in sequence before taking a Standard scenario eat decision. Internal actions are used to perform analysis of The first chart in Figure 8 describes the behavior of Adam options and to execute queries to the knowledge base that during the standard scenario (eat the Schnitzel). The first modify the internal state of the possible goal. column shows Adam’s current plan. The combination of high hunger drive (see Figure ), the perception of a Calibration and Simulation Schnitzel and the memorized satisfaction for eating The modules in SiMA encapsulate functionalities of the Schnitzel, make Adam initially follow the plan to EAT. human mind and are developed independently, following a After the Schnitzel was consumed, Adam switches to the black-box approach. Meaningful integration tests for these BEAT plan, as it modules require a level of knowledge about module fits the new interaction, which is not available, due to the high number perception (Bodo; of modules, parameterization options and their functional no Schnitzel) and structure. Therefore we keep integration testing to a new drive state. minimum in favor of system testing, using exemplary cases. Figure visualizes Calibration is performed in various steps on each scenario. the changes in First the environmental situation (Adam, Bodo, and the Adams drive state Figure 9: Drives: hunger in green Schnitzel) is modelled as the most basic layer of calibration. in detail. Adam (aggressive) and red (libidinous). Next, the drive situation is modelled and memories are starts with high created to match memorized actions to drives, according to hunger. While he eats, the hunger drops, since eating the the simulation case description. Where needed, the defense Schnitzel changes Adam’s body state which the drives mechanisms are modelled and harmonized with the drive represent. In time, the hunger subsided below the sexual situation. Lastly, the acts are modelled and associated to the drives, which started out low but steadily increase. The memorized actions. Each step could, and often did, require stamina drives (Figure 9 in blue and cyan) represent Adam’s previous steps to recalibrate to allow modeling according to need for relaxation and changed in response to Adams the description. This resulted in a calibration strategy exhaustion while approaching the Schnitzel (first two peaks) similar to a waterfall model with feedback. and Bodo (third peak). Alternative scenario beat Simulation Results This scenario differs from the standard scenario in its initial drive state. Adam starts with higher, faster increasing As mentioned, we validate our model via test scenarios in sexual drives and low hunger. The BEAT plan is memorized the MASON simulation framework. This chapter with the highest satisfaction for the sexual drives and is summarizes the results of these simulations. We compare associated with the current emotional state (see Figure 10). the agent’s behavior and internal state to the expectations Beating reduces the anger and causes a short peak of joy. 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 Figure 10: Emotional state in the beating scenario. indicate sight ranges, the green Agent will be referred to as Alternative scenarios give and share Adam, the red agent as Bodo and the round shape between These scenarios use a defense mechanism to alter Adam’s them as Viennese Schnitzel. 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 Figure 8: Behavior sequences in the simulation case. standard scenario, with the hunger drives dominating. A super-ego rule is introduced to create a conflict between 2 Plans are written in capital letters to distinct them from actions. high hunger drives and perceiving Bodo with the Schnitzel. 519 The conflict is resolved by subliming the action eat, which Bargh, J.A., & Chartrand, T.L., (1999). The unbearable is chosen due to its high memorized satisfaction for the automaticity of being. American psychologist, 54(7): 462- hunger drive (see description of standard scenario eat), with 479. the action give or share. In both cases Adam will follow the Braitenberg, V., (1984). Vehicles: Experiments in synthetic corresponding plan, as the sublimed action now promises psychology, MIT Press, Cambridge. the highest satisfaction for the hunger drive. Bruckner, D., Gelbard, G., Schaat, S., & Wendt, A. (2013). Validation of cognitive architectures by use cases. Lessons learned Proceedings of Industrial Electronics (ISIE), IEEE The calibration of drives makes for a good point of International Symposium on. departure for calibrating the SiMA model. One possible Damasio, A.R., (2003). Looking for Spinoza: Joy, Sorrow, problem at this stage is that the drive situation also and the Feeling Brain. Harvest Books. influences the perception. Calibrating the perceived reality Duch, W., Oentaryo, R. J., & Pasquier, M. (2008). is normally done straight forward by directly specifying the Cognitive Architectures: Where do we go from here?, simulated world, but extreme drive states influence the Frontiers in Artifcial Intelligence and Applications, vol selection of memories for perception, since the agent also 171, pp. 122-136. activates memories that are that satisfy current drives. After Dietrich, D., et al. (2014). Naturwissenschaftliches, modelling the base behavior, the defense mechanism can be psychoanalytisches Modell der Psyche. Technical Report used to fine tune the agent. They can alter the influence II, Vienna University of Technology, Institute of other primary process modules have on the decision. In the Computer Technology, Austria. current simulations, the defense was only used to model Faghihi, U., & Franklin, S. (2012). The Lida model as a alternative behavior in situations that match certain criteria foundational architecture for AGI. In Theoretical (e.g. sharing when Bodo is next to the food source). More Foundations of Artificial General Intelligence (pp. 103– elaborated uses of this tool are thinkable as the defense is 121). Springer. capable of influencing the impact other primary process Laird, J. E., Congdon, C. B., Coulter, K. J., Derbinsky, N., modules have on the decision (e.g. fine tuning the drive & Xu, J. (2011). The Soar user’s manual version 9.3.1. state via super-ego rules). This ability to influence, and even Technical report. Computer Science and Engineering alter, the results of other modules, also makes the defense Department, University of Michigan. harder to calibrate, since changes in this module may affect Langley, P., Choi, D., & Trivedi, N. (2011). Icarus user’s many others and vice versa. manual. Technical report, Institute for the Study of Learning and Expertise 2164 Staunton Court, Palo Alto. Conclusion and Future Work Langley, P., Laird, J. E., & Rogers, S. (2009). Cognitive Architectures: Research Issues and Challenges, Cognitive We showed how a narrative exemplary case that Systems, 10, 141-160. exemplifies assumptions and requirements for a cognitive Schaat, S., Doblhammer, K., Wendt, A., Gelbard, F., Herret, architecture is structured into a deterministic description and L., & Bruckner, D., (2013). A Psychoanalytically-Inspired test plan. The resulting model is evaluated in simulations, Motivational and Emotional System for Autonomous where we showed how the interplay of various decision Agents. Proceedings of the 39th Annual Conference of the factors, such as drives, emotions, social norms, results in an IEEE Industrial Electronics Society. adapted decision for an agent’s internal and external state. Schaat, S., Wendt, A., Jakubec, M., et al. (2014). ARS: An Using a functional model we emphasized on how behavioral AGI Agent Architecture. In B. Goertzel, L. Orseau, J. change is generated without changing the model. Snaider (Eds.). Artificial General Intelligence. The presented evaluation using simulation cases can be Proceedings of the 7th International Conference, AGI (pp. regarded as a calibration that validates the transformation of 155-164). New York, USA: Springer. the SiMA model into a software implementation. However, Schaat, S., & Dietrich, D., (2014). Case-Driven Agent- such calibration is only a first step to demonstrate the Based Simulation for the Development and Evaluation of model’s consistency and plausibility, with the next step Cognitive Architectures. Proceedings of the 26th Benelux being to observe if unspecified parameters would generate Conference on Artificial Intelligence (pp. 73-80). comprehensible results without model adaptions. Solms, M. (2009). What is the ‘Mind’? A Neuro- In the end such methodology is a premise to apply the Psychoanalytical Approach. In Dietrich D., Fodor G., model in a specific domain, e.g. as a decision support tool Zucker G., & Bruckner D. (Eds.). Simulating the Mind for marketing strategies (as currently approached), where (pp. 115-122). Wien, Austria: Springer. we will additionally test our model against empirical data. Vernon, D., von Hofsten, C., & Fadiga, L. (2011). A Roadmap for Cognitive Development in Humanoid References Robots. Springer, Berlin. Anderson, J. R., Bothell, D., Byrne, M. D., Douglass, S., Lebiere, C., & Qin, Y. (2004). An integrated theory of the mind. Psychological Review, 111(4):1036–1060. 520