Agent Architecture: An Overview ‘Scanned with CamScannerIntroduction The advancement of the Intemet technology has increased the need for distributed, concurrent, heterogeneous and dynamic application systems. Agent technology is a new paradigm suitable for developing such systems that situates and operates in a dynamic and heterogeneous environment. What exactly is an agent? To date, there is no widely accepted definition of what an agent is. In this study, an agent is referred to as an autonomous software entity that is situated in some environment where it can monitor and response to changes proactively or reactively by itself or through communication with other agents to persistently achieve certain goal/task on hehalf of user or other agents (Wooldridge, 2009). An agent possesses certain distinct characteristics such as (Wooldridge & Jennings, 1995). * Autonomous: the ability to operate without the direct intervention of human and control over its internal state. e Social: the ability to interact with human and other agents. » Reactive: the ability to perceive changes in the environment and response to it ina timely fashion. * Proactive: the ability to show goal directed behavior. Other characteristics that an agent might have include mobility, benevolent, trustworthiness, rationality, and learning capability. Mobility is the ability to travel between different hosts in a computer network. Benevolent is the characteristic agent will always perform what it is asked to do. ‘Scanned with CamScannerTrustworthiness is the characteristic agent will not deliberately communicate false information. Rationality is the characteristic agent will always never to prevent its goals being achieved. Learning capability is the ability to adapt itself to fit its environment and to the desires of its users. In a complex system, an agent may not exist alone in an environment as they may be multiple agents that are situated in the same environment. Multi-agent system is the study of systems that are made up of multiple heterogeneous software entity (agents) that interact with each other (Wei B, 1999; Shoham & Leyton-Brown, 2008). In a multi-agent system environment, agents may have common or conflicting goal to be achieved (Yu et al., 2010; Durfee & Rosenschein, 1994), The interaction between agents can happen directly or indirectly. Direct communication is achieved through channel such as message passing whilst indirect communication is achieved through affecting the environment and sense by other agents (Genesereth & Ketchpel, 1994; Maes, 1997). Normally, agents that have common goal in a multi-agent system will cooperate in order to achieve the goal (Doran et al., 1997; Pozna et al., 2011). In the case of agents with conflicting goals, the agents will compete against each other to obtain resources for personal goal attainment (Leyton-Brown , 2003). In order for agents to cooperate and coordinate in achieving their goals, agents need to reason about when and what to do under certain circumstances. The foundation of the agent reasoning mechanism lies in the component called agent architecture. Agent architecture is the blueprint for building an agent just like a class in object- oriented programming. Wooldridge referred to agent architecture as software architecture that is intended to support decision making process (Wooldridge, 2001). Maes described agent architecture as architecture that encompasses techniques and algorithms to support decomposing set of components and how these components interact (Maes, 1991). Agent architecture is the building block for creating an agent much like creating an object in a class. The agent architecture is the brain of the agent as it determines how the knowledge/information is represented in the agent. It also determines the action the agent should take based on its underlying reasoning/interpretation mechanism. Thus, different architectures used different representation approaches for their reasoning mechanism to solve a variety of problems. These architectures can be broadly categorized into three groups, the classical architecture, the cognitive architecture and the semantic agent architecture. The classical architectures include logic-based architecture, reactive architecture, BDI architecture, and hybrid architecture. The logic-based architecture is an agent architecture that uses symbolic Tepresentation for reasoning. The reactive agent architecture is a direct stimulus-response agent architecture. On the other hand, the BDI architecture is a deliberative agent architecture based on mental states characteristic such as belief, desire, and intention. The layered architecture is the hybrid of reactive and deliberative agent architecture. The cognitive architecture is based on cognitive sciences and the semantic agent architecture utilizes semantic technology. ‘Scanned with CamScannerThe remainder of this paper describes the various agent architectures that can be used to build agent and multi-agent system, Section 2 discusses the logic-based architecture. The reactive agent and BDI architecture are discussed in Section 3 and 4 respectively. Section 5 describes the layered architectures whilst Sections 6 and 7 present an overview of cognitive architecture and semantic agent architecture, Finally, we conclude our discussed on agent architectures in Section 8. Logic-Based Architecture Logic-based architecture also known as the symbolic-based or deliberative architecture is one the earliest agent architecture that rests on the physical-symbol systems hypothesis (Newell & Simon, 1976). This classical architecture is based on the traditional artificial symbolic approach by representing and modeling the environment and the agent behavior with symbolic representation. Thus, the agent behavior is based on the manipulation of the symbolic representation. Agent's role in this classical architecture may also be considered as theorem provers (Shardlow, 1990). The syntactical manipulation of the symbolic representation is the process of logical deduction or theorem proving. As an instance of theorem proving, the agent specifications outlines how the agent behaves, how the goals are generated and what action the agent can take to satisfy these goals. An example of logic-based architecture formalism is as follows: « Assume that the environment is described by sentences in L and the knowledge base that contains all the information regarding the environment KB = P(L) where P(L) is the set of possible environments. © For each moment of the time t, an agent’s internal state is represented by KB = {KB,, KB,, KB... KB,} where KBE KB. «The possible environment states are represented by $ = (1,52, »/. « Anagent's reasoning mechanism is modeled by a set of deduction rules, p which are the rules of inference. © Anagent perception functions as see:S ->P. © The agent's internal state is updated by a perception function where next:KB x P ->KB. * Thus, agent can choose an action from a set A = {@,, a, .../, action: KB ->A which is defined in terms of deduction rules. The outcome of an agent's actions is drawn via the function do where do:A x $ ->S. ‘© The decision making process is modeled through the rules of inference p, if a do:A can be derived, the A is returned as an action to be best performed, else if do:A cannot be derived, a special null action is returned. ‘Vacuum cleaning example in (Russell & Norvig, 1995) illustrates the idea of logic-based architecture based on the specification above. The programmer has to encode the inference rules p in a Way that enables the agent to decide what to do. Examples of this kind of classical agent architecture ‘Scanned with CamScannerapproach include classical planning agent such as STRIPS (Fikes & Nilsson, 1971), IPEM (Ambros- Ingerson & Steel, 1988), Autodrive (Wood, 1993), Softbots (Etzioni et al., 1994), Phoenix systems (Cohen et al., 1989), IRMA (Bratman et al., 1968), HOMER (Vere & Bickmore, 1990), and GRATE (Jennings, 1993). BDI architecture is also considered as the subset of logic-based architecture, However, due to its popularity and wide adoption of the architecture, the discussion on this particular architecture is detailed in Section 4. Although, the simplicity and elegance of logical semantics of the logic based architecture is. attractive, there are several problems associated with this approach. Firstly, the transduction problem implies the problem of translating modeling into symbolic representation. It is difficult to translate and model the environment’s information into symbolic representation accurately for computation process especially complex environment. Secondly, it is also difficult to represent information in a symbolic form that is suitable for the agents to reason with and in a time constrained environment. Finally, the transformation of percepts input may not be accurate enough to describe the environment itself due to certain faults such as sensor error, reasoning error and etc. It is very difficult or sometimes impossible to pul down all the rules for the situation that will be encountered by the agent in a complex environment since the deduction process is based on set of inference rules. The assumption in calculative rationality where the world does not change in a significant way while the agent is dekberating is not realistic. Assume that on time t,, agent tries to reason an optimal action for that particular time. However, the reasoning result may only be available at time t; where the environment has already changed so much so that the optimal action for time t, may not be an optimal action for time ty, Thus, due to the computational complexity of theorem proving over this approach, it is not appropriate for time constrained domain. Building agent in logic-based approach is viewed as a deduction process. An agent is encoded asa logical theory by using specification and the process of selecting the action is through deduction process that reduces the problem to a solution such as in theorem proving. An improvement version logic based approach has been carried out in (Amir & Maynard-Reid, 2004; Amir & Maynard-Reid, 2000). In (Amir & Maynard-Reid, 2004) a logic-based Al architecture is implemented on Brooks’ Subsumption architecture. 3a the implementation of this architecture, different layers of control is axiomatized in First-Order Logic (FOL), thus, independent theorem provers are used to derive each layer’s output given its input, This architecture proved the versatility of the theorem provers wiich allow them to realize complex tasks, while keeping individual theories simple (Amir & Maynard- Reid, 2000). Reactive Architecture Reactive agent architecture is based on the direct mapping of situation to action. It is céfferent from the logic-based architecture where no central symbolic world model and complex symbolic reasoning are used. Agent responses to changes in the environment in a stimulus-response based. The reactive ‘Scanned with CamScannerarchitecture is realized through a set of sensors and effectors, where perceptual input is mapped to the effectors to changes in the environment. Brook's subsumption architecture is known as the est pure reactive architecture (Brooks, 1986). This architecture was developed by Brook who has critiqued on many of the drawbacks in logic-based architecture. Figure 1 illustrates an example of reactive architecture. The figure shows that each of the percept situation is mapped into an action which specifically responses to the percept situation. Perception Environment Figure 1; Reactive Architecture The key idea of subsumption architecture is that intelligent behaviour can be generated without explicit representations and abstract reasoning with symbolic AI technique (Brooks, 1991a; Brooks, 1991b). Intelligence is an emergent property of ceriain complex systems. Subsumption architecture is implemented in finite state machines with different layers connected to sensors that perceive the environment changes and map the action to be performed (Brooks, 1986). A set of task- accomplishing behaviour are used in the decision making process. Fach of the behaviour can be thought of as an individual function which maps changes in the environment with an action. Multiple behaviours that can be fired simultaneously is another characteristic of subsumption architecture. The subsumption architecture hierarchical structure represents different behaviours. The lowest layer in the hierarchy has the highest priority. Higher layer represent more abstract behaviour than the lower layer in the hierarchy. Complex behaviour is achieved through the combination of these behaviours. Figure 2 shows action selection in the layered architecture. In this layered architecture, the lower the layer the higher the priority. The lower layer will be the primitive behaviour and higher layer wilt Tepresent a more abstract behaviour. ‘Scanned with CamScannerFigure 2: Action Selections in Layered Architecture ‘The subsumption architecture is implemented in Steel’s study (Steels, 1990) for the mission to a distant planet to collect sample of rocks and minerals. A near-optimal performance can be obtained through simple adjustment and the solution is cheap in computing power and it is also robust. Chapman came up with similar approach to Brook’s work in (Chapman & Agre, 1986) and is referred to as the new abstract reasoning, This approach is used in the celebrated PENGI system which simulated a computer game with central character control that can accomplish routine work with little variation (Agre & Chapman, 1987). Figure 3 shows a Pengo game in progress (Agre & Chapman, 1987),PENGI is a program written to play an arcade game called Pengo which is made up of a 2-Dmaze with unit-sized ice blocks. In this game, PENGI is programmed to move the Penguin in the game to avoid the bees attack and block slide to survive. Figure 3: Pengo Game in Progress ‘Scanned with CamScannerAnother similar approach is the situated automata paradigm by Rosenschein and Kaelbling (Kaelbling, 1991; Kaelbling & Rosenschein, 1990; Rosenschein, 1985; Rosenschein & Kaelbling, 1986). In this approach, the agent is specified in declarative terms such as beefs and goals which are then compiled into a digital machine which satisfies this declarative specification. The digital machine can operate in a provably time-bounded fashion. Although the declarative terms are in beliefs and goals, the digital machine does not use any of the symbolic representation and hence, no symbolic reasoning actually occurred. The approach is also considered as reactive architecture as no symbolic Teasoning actually occurred because the declarative terms has been compiled to a digital machine that reacts based on stimulus response. Another reactive architecture is introduced by Maes called the Agent Network Architecture. In this architecture, the agent acts as a set of competence modules (Maes, 1991; Maes, 1989; Maes, 1990). These modules loosely resemble the behaviours of Brook’s subsumption architecture. Pre and post-conditions determine the activation of each module based on an activation level. This activation level is stated in real value to indicate the relevancy of the module in a particular situation. Agent network architecture is somewhat similar to the neural network architecture with different meaning of the node context. One of the advantages of reactive architecture is that it is less complicated to design and implement than logic-based architecture. An agent’s behaviour is computationally tractable. The robustness of reactive architecture against failure is another advantage, Complex behaviours can be achieved from the interaction of simple ones. The disadvantages of reactive architecture include (1) insufficient information about agent's current state to determine an activation action due to modeling of environment available, (2) the processing of the local information limits the planning capabilities in long term or bigger picture and hence, learning is difficult to be achieved, (3) emergent behavior which is not yet fully understood making it even more intricate to engineer. Therefore, it is difficult to build task-specific agents and one of the solutions is to evolve the agents to perform certain tasks (Togelius, 2003). The work in this domain is referred to as artificial life. Belief-Desire-Intention (BDI) Architecture The BDI architecture is based on practical reasoning by Bratman’s philosophical emphasis on intentional stance (Bratman, 1987). Practical reasoning is reasoning toward actions - the process of figuring out what to do. This is different from the theoretical reasoning process as it derives knowledge or reaches conclusions by using one’s beliefs and knowledge. Human practical reasoning involves two activities namely deliberation and means-end reasoning. Deliberation decides what state of affairs needs to be achieved while means-end reasoning decides how to achieve these states of affairs. In BDI architecture, agent consists of three logic components referred as mental states/mental attitudes namely beliefs, desires and intentions. Beliefs are the set of information an agent has about the world. Desires are the agent's motivation or possible options to carry out the actions. Intentions ‘Scanned with CamScannerare the agent’s commitments towards its desires and beliefs. Intentions are key component in practical reasoning. They describe states of affairs that the agent has commitied to bringing about and as a result they are action-inducing. Forming the intentions is critical to an agent's success. BDI architecture probably is the most popular architecture (Rao & Georgeff, 1991) and Practical Reasoning System (PRS) is one of the well known BDI architectures (Georgeff & Lansky, 1986). PRS is a framework for building real-time reasoning systems that can perform complex tasks in dynamic environments. PRS used procedural knowledge representation in describing how to perform a set of actions in order to achieve goal. This architecture is based on four key data structures: beliefs, desires, intentions and plans, and an interpreter (see Figure 4). Figure 4 shows the four key data structures of BDI architecture. In the PRS system, beliefs represent the information an agent has about its environment. Desires represent the tasks allocated to the agent corresponding to the goals that should be accomplished by agent. Intentions represent the agent's commitment towards the goals. Finally, plans specify some courses of action for the agent in order to achieve its intentions. The plans in the plan library are pre-compiled plan rather than instantaneously generated. The agent interpreter is responsible for updating beliefs from observations made from the environment, generating new desires (tasks) on the basis of new beliefs, and selecting from the subset of currently active desires to act as intentions. Lasily, the interpreter must select an action to perform the agent's current intentions and procedural knowledge. Figure 4: Practical Reasoning Systems ‘Scanned with CamScannerMany agent frameworks have been implemented in BDI architecture including: + JAM, a hybrid intelligent agent architecture that draws upon the theories and ideas of the PRS, Structured Circuit Semantics (SCS), and Act Plan Interlingua (Huber, 1999), « JACK, a commercial platform for developing industrial and research purpose multi-agent application in Java (Howden et al., 2001). e dMARS is a platform for intelligent software agents that makes use of the belief-desire— intention software model (BDI) for building complex, distributed, time-critical systems in C+ (a'Inverno et al., 1998). The advantages of BDI architecture are that the design of the architecture is clear and intuitive. The functional decomposition of the agent subsystem is clear and the BDI logic has formal logic properties that can be studied. However, with BDI architecture the question of how to efficiently implement the functionality in subsystem is not clear and so the agents need to achieve a balance between commitment (Rao & Georgeff, 1991) and reconsideration (Wooldridge & Parsons, 1998). If an agent did not stop to reconsider, it might be trying to achieve intention which is not achievable or no longer valid. If an agent reconsiders often, it might face the risk of not achieving them due to insufficient time working on the task. Layered (Hybrid) Architecture Layered (hybrid) architecture is an agent architecture which allows both reactive and deliberate agent behavior. Layered architecture combines both the advantages of reactive and logic-based architecture and at the same time alleviates the problems in both architectures. Subsystems are decomposed into a Jayer of hierarchical structure to deal with different behaviours. There are two types of interaction that flow between the layer namely horizontal and vertical. In the horizontal layer architecture, each layer is directly connected to the sensory input and action output (see Figure 5). Each layer is like an agent mapping the input to the action to be performed. TouringMachine is an example of horizontally layered agent architecture. The TouringMachine agent architecture consists of three activity-producing layers: a reactive layer R, a planning layer P, and a modeling layer M (Ferguson, 1992). These three layers operate concurrently and independently in mapping the perception into action (see Figure 6). Each of the layers has its own internal computation processing mechanism. ‘Scanned with CamScannerLayern Layern-1 Layer n-2 uony Perception Figure 5: Horizontal Layer Architecture Sensory Action Input Modelling Layer (M) h Output Perception 7 Action Subsystem Planning Layer (P) Subsystem Reactive Layer (R) ‘Control Framework Figure 6: The TouringMachine Agent Control Architecture ‘The advantage of horizontal layer architecture is that only n layers are required for mapping to n different types of behaviours. However, a mediator function is used to control the inconsistent actions between layer interactions. Another complexity is the large number of possible interactions between horizontal layers-m (where m is the number of actions per layer). Vertical layer architecture eliminates some of these issues as the sensory input and action output are each dealt with by at most one layer each (creating no inconsistent action suggestions). There are two types of vertical layered architectures namely one-pass and two-pass control architectures. In one-pass architecture, control flows from the initial layer that gets data from sensors to the final layer that generates action output (see Figure 7), In two-pass architecture, data flows up the sequence of layers and control then flows back down (see Figure 8). ‘Scanned with CamScannerAction . Layern ' ¥ Layern , wert » bayer n-a tae x layern-2 4 " ~ ¥ >-——__ ——-— * - + a > t }—+ F#— a Layer2 Layer 2 4 + Layer 1 Layer —_— - Perception Perception Action Figure 7; Vertical layer architecture: one-pass Figure 8: Vertical layer architecture: two-pass InteRRaP is an example of vertically two-pass agent architecture (Muller & Pischel, 1993). InterRRaP comprises of three control layers namely behavior layer, local planning layer and cooperative planning layer (see Figure 9), The behaviour layer is the lowest layer in InteRRap which Tesponses reactively towards the world model. The local planning is the middle layer that uses planning knowledge for the routine planning to achieve the agent's goal. Finally, the cooperative planning is the highest layer in InteRRaP that deals with the social interaction. The main difference between InteRRaP and TouringMachine is the interaction between layers. There are two types of flow control in InteRRaP which are the bottom up and top down. The bottom up activation deals with the Jower layer which passes control to the higher layer when it cannot process the current situation. Top down execution deals with the higher layer which uses the action execution of the lower layer to achieve goals and tasks. There are two general functions that are implemented in each layer known as the situation recognition and goal activation function and planning and scheduling function. The situation recognition and goal activation function is responsible to map the knowledge base and current goals to a new set of goals. The planning and scheduling function is responsible for selecting which plans to execute, based on the current plans, goals and knowledge base of that layer. ‘Scanned with CamScannerFigure 9: InteRRap Architecture The main advantage of vertical layered architecture is the interaction between layers is reduced significantly to m’(n—1). The main disadvantage is that the architecture depends on its robustness, so if one layer fails, the entire system fails. Cognitive Architecture As stated in (Langley, 2004), there are several ways of constructing intelligent agents which are the software engineering approach, the multi-agent approach and the cognitive architecture approach. The cognitive architecture is based on the stream of cognitive science. The study of cognitive science focuses on the human cognition and psychology. The origin of cognitive architecture started with a specific class of architecture known as production systems (Neches et al., 1987; Newell, 1973) and evolves over time. ACT is one of the earliest cognitive architecture build for modeling human behavior (Anderson, 1976). The construction of intelligent agent using cognitive architecture focuses on the cognition part, the simulation and modeling of human behaviour. The cognitive architecture approach is said to be different from multi-agent approach in the following manners (Langley et al., 2009): cognitive architecture comes with a programming formalism to encode knowledge and associates it with ils interpreter, + it has strong assumptions on the representation of knowledge and the processes that operate on them, ‘Scanned with CamScannerit assumes a modular representation of knowledge, © it offers intelligent behavior at the systems level, rather than at the level of component methods designed for specialized tasks, and + it provides a unified approach in which a common set of representations and mechanisms reduces the need for such careful crafting. Cognitive architectures are used to construct intelligent system/agent that models human performance (Newell, 1990; Meyer & Kieras, 1997). Thisunderlying cognitive architecture should typically possess the characteristics of (Langley et al., 2009): «short and long-term memories for storing the agent’s beliefs, goals, and knowledge, representation of memories and their organization, and + functional processes that operate on these structures. Memory and learning are the two important key components in developing cognitive architecture. The combination of these two components form a taxonomy of three main categories of cognitive agent architecture namely symbolic, emergent, and hybrid models (Duch et al., 2008) (see Figure 10). Symbolic architecture is a classical AI top-down analytical approach that focuses on information processing by using high-level symbols or declarative knowledge. Examples of cognitive symbolic architecture are SOAR (Laird et al., 1987; Newell, 1990), EPIC (Meyer & Kieras, 1997), ICARUS (Langley, 2005). SOAR is a platform created to demonstrate general intelligence behaviour. The platform is based on symbolic representation which utilizes operator associated to problem space in production rules. This approach is referred as procedural long term knowledge. The platform has been used to develop many of the applications including expert system, intelligent control, and human. behaviour interaction simulation. SOAR platform can also be viewed as a theory of general intelligence or as theory of human cognition. The architecture is based on human problem solving and unified theory of cognition, The theory of the architecture is that the unification of different theories or overlap theories without conflict can produce general intelligence behaviour with appropriate learning mechanism. A bottom-up approaches which differentiates it from the symbolic architecture is adopted in emergent architecture whereby low-level activation signals flowing through a network consisting of numerous processing units is used. IBCA (O'Reilly et al., 1999), Cortonis (Hecht-Nielsen, 2007), NuPIC (Hawkins & Blakeslee, 2004) are a few of the systems under the category of emergent architecture. IBCA is based on the working memory as contro¥ed processing on how the components in brain dynamically interact with each other to bring about the cognition function. The emergent property of the dynamic interactions of the components contributes the working model of IBCA. There are several important components in IBCA and each has own specific function namely the posterior perceptual and motor cortex (PMC) which are responsible for the sensor and motor processing based on inference and generalization. The prefrontal cortex (PFC) is responsible for ‘Scanned with CamScannerdynamic and active memory and hippocampus (HCMP) is in charge of the rapid leaning of arbitrary information. Hybrid architecture is the combination of the symbolic and emergent paradigms. Examples in this category of hybrid architectures are ACT-R (Anderson & Lebiere, 2003), CLARION (Sun & Alexandre, 1997; Sun et al., 2001) and LIOA (Franklin, 2006). ACT-R is one of the earliest developed cognitive architecture which is based on the modeling of human behaviour. ACT-R is organized into of a set of modules namely, sensory module, motor module, intentional module and declarative module. Each module has a temporary storage which holds the short term memory, A long-term memory is achieved through the combination of these modules. The action selection is based on the utility calculation, in which the highest utility actions are selected.Figure 10 shows the taxonomy of cognitive architecture based on the two important design properties which are memory and learning. Symbolic Emergent Hybrid Memory i Memory Memory ‘* Rule-basedmemory © Globalist memory © Localist-distributed _ © Graph-based memory _* Symbolic-connectionist : 1 <= I Learning Learning Learning © Inductive leaming © Associative leaming ‘* Bottom-up leaming © Analytical leaming © Competitive leaning * Top-down learning Figure 10: Taxonomy of Cognitive Architectures (Duch et a., 2008) Hybrid architecture is the combination of the symbolic and emergent paradigms. Examples in this category of hybrid architectures are ACT-R (Anderson & Lebiere, 2003), CLARION (Sun & Alexandre, 1997; Sun et al., 2001) and LIOA (Franklin, 2006). ACT-R is one of the earliest developed cognitive architecture which is based on the modeling of human behaviour. ACT-R is organized into of a set of modules namely, sensory module, motor module, intentional module and declarative module. Each module has a temporary storage which holds the short term memory. A long-term memory is achieved through the combination of these modules. The action selection is based on the utility calculation, in which the highest utity actions are selected. Figure 10 shows the taxonomy of cognitive architecture based on the two important design properties which are memory and learning. ‘Scanned with CamScannerConclusion Agent architecture is the key component to constructing agent. It acts as the brain and heart of the agent to reason and perform action based on its knowledge base. This paper provides the state of the art of agent architecture. It outlined four main agent architectures: logic, reactive, BDI and layered architecture. Apart from these four, the cognitive architecture which is based on the human behaviour is another architecture that is widely adopted in building multi-agent system. Many of the foundations of cognitive architecture rest on the theories in cognitive sciences domain. The adoption semantic web technology to enhance the cognitive architecture in analyzing human behaviour will be a very interesting area to explore. Finally, the semantic technology adoption into agent architecture is also another interesting topic that should be studied as well. For future work, we would like to draw on these preliminary reviews to study the potential of a new agent architecture that uses semantic web technology to enable agent to process the information in a more meaningful way that will in tum improve its decision making which closely mimics the human reasoning process. ‘Scanned with CamScannerAbstract Artificial intelligence, and in particular machine learning, is a fast-emerging field. Research on artificial intelligence focuses mainly on image-, text- and voice-based applications, leading to breakthrough developments in self-driving cars, voice recognition algorithms and recommendation systems. In this article, we present the research of an alternative graph- based machine learning system that deals with three-dimensional space, which is more structured and combinatorial than images, text or voice. Specifically, we present a function- driven deep learning approach to generate conceptual design. We trained and used deep neural networks to evaluate existing designs encoded as graphs, extract significant building blocks as subgraphs and merge them into new compositions. Finally, we explored the application of generative adversarial networks to ‘Scanned with CamScannergenerate entirely new and unique designs. ‘Scanned with CamScanner—»> Evaluating Designs Design > Data NN : Discovering : J I—> Building -O : : Blocks 00 pur taver | IDDEN LAYERS | ourpur Laver Figure 1. Abstract representation of deep neural networks (DNNs). Data are fed into the input layer, which is processed in the hidden layers (there can be any number of hidden layers), and a result is produced in the output layer. DNNs are classified as supervised or unsupervised. Supervised DNNs are trained with labelled data, that is, data that contain information on type, attributes or scores. In contrast, unsupervised BNNs are fed with raw data without any added information about the ‘Scanned with CamScannercharacteristics of the data. In our research, we used both systems; for one study, we used labelled design data of various nouses (represented as graphs with node and edge attributes) with functional scores and hence utilized a supervised DINN. In another study, we used labelled design data, but without the functional scores, which can be considered an unsupervised DNN. The training set for our DNN consisted of graph representations of architectural designs, since this mode of representation can capture structused relationships between system components much better than continuous media formats, for example, images. Graphs are mathematical structures comprising nodes (denoting system components) and edges (denoting relationships between nodes), and they lend themselves to mathematical and algorithmic analysis at the level of system ‘Scanned with CamScannercomponents. Graph theory has been used as a powerful representational and analytical tool in diverse fields such as social network analysis, transportation network analysis, biological network modelling and Internet engineering. In our study, node and edge attributes capture interesting physical characteristics of architectural design components, for example, area and volume. In contrast, graph attributes characterize the overall design functions, for example, liveability and sleepability. These are exposed to graph-based DNNs, which attempt to find patterns occurring in architectural design samples. Such an analysis is hard for pixel- based images, which do not carry any explicit semantics of architectural designs. ‘Scanned with CamScanner2. The Contemporary Concrete Mix Design and Machine Learning Techniques 2.1. Concrete Mix Design in European Corporate Practice The primary goal of concrete mix design is to estimate the proper quantitative composition and proportion of concrete mixture components. We should use a composition which allows us to achieve the best possible concrete performance. Concrete performance is characterised by several features, from which the most significant are compressive strength and durability. Both concrete strength and durability should play an essential role in the concrete mix design. The issue of durability is essential in the case of an aggressive environment [17,18,19,20,21,22]. Based on industry ‘Scanned with CamScannerexperience, we found out that, in European corporate engineering practice, there are a few most used methods for designing a concrete mix. These methods are the Bukowski method, the Eyman and Klaus method, and the Paszkowski method. All the solutions mentioned above are derived from the so-called “Three Equations Method”, or Bolomey Method, which is a mixed experimental- analytical procedure [10,23,24]. It means that collected laboratory data should confirm that mathematical approach. We calculate a volume of required components by analytical measures and validate the results by destructive laboratory testing. In this method, we use a fundamental equation of strength, consistency, and tightness to determine the three searched values, as follows, the amount of aggregate, cement, and water, expressed in kilograms per cubic meter. ‘Scanned with CamScannerThe first equation is the compressive strength equation or Bolomey formula (Equation (1)), which expresses the experimentally determined dependence of the compressive strength of hardened concrete on the grade of cement used, the type of aggregate used, and the water- cement ratio characterising the cement paste [23,24]. In this method, the concrete grade is assumed as input data. Cc tem = Ai GF +0, 5) [MPa], (1) where, fom is a medium compressive strength of concrete, expressed in kilograms. The value A; 2 means coefficients, depending on the grade of cement and the type of aggregate; C is an amount of cement in 1 m? of concrete, expressed in kilograms; and W corresponds to the amount of water in 1 ‘Scanned with CamScannernn m? of concrete, expressed in kilograms. A second consistency Equation (2), is included in the water demand formula necessary to make a concrete mix with the required consistency. W =C-w. +K - wx [dm], (2) where W is the amount of water in 1 m3 of concrete, expressed in kilograms; C corresponds to the amount of cement in 1 m of concrete, expressed in kilograms; K means the amount of aggregate in 1 m? of concrete, expressed in kilograms; w, is the cement water demand index in dm? per kilogram; and w, is the aggregate water demand index in dm? per kilogram. The water-tightness of concrete Equation (3) is included in the simple volume formula, which indicates that a watertight concrete mix is obtained if the sum of the volume of the individual components is equal to ‘Scanned with CamScannerthe volume of the concrete mix. £ + * 4 w = 1000 [dm*], (3) Pc Pk where W is the amount of water in 1 m? of concrete, expressed in kilograms, C corresponds to the amount of cement in 1 m/® of concrete, expressed in kilograms, K means the amount of aggregate in 1 m° of concrete, expressed in kilograms; p, is the cement density in kilograms per dm*; and pr is the aggregate density in kilograms per dm. The system of equations presented above, with three unknowns variables, allows for calculating the sought amounts of cement (C), aggregate (K), and water (W) in one cubic meter of concrete mix. The system is valid, assuming that there are no air bubbles in the concrete. Another method used in the construction industry is “the ‘Scanned with CamScannerdouble coating method” [25]. The methods above are ones that are used to determine the quantitative composition of the concrete mix. However, the actual process of creating a concrete mix is much broader, including the following steps: The first step is to determine the data needed to design the mix, such as the purpose of the concrete use, the compressive strength of the concrete, and the consistency of the concrete mix. Next, the qualitative characteristics of the components should be determined, namely the type and class of cement and the type and granularity of the aggregates. Subsequent steps include an examination of the properties of the adopted ingredients; a check of their compliance with the standard requirements; determining the characteristics of the components that will be needed to ‘Scanned with CamScannera determine the composition of the concrete mix; and a projection of the aggregate pile. The successive step is the actual adoption of the design method and a calculation per unit of volume. The final stage is to make a trial sample and examine both the concrete mix and the hardened concrete with design assumptions [26]. ‘Scanned with CamScanner2 KQML The Knowledge Query and Manipulation Language (KQML) is a language and an associated protocol to support the high level communication among intelligent agents. It can he used as a language for an application program to interact with an intelligent system or for two or more intelligent systems to interact cooperatively in problem solving. We argue that KQML should he defined as more than a language with a syntax and semantics, but must also include a protocol which governs the use of the language (c.g., a pragmatic component ) 2.1 Design Issues and Assumption Architectural assumptions. Agents will typically be separate processes which may be running in the same address space or on separate machines. ‘The machines will he accessible via the internet, We need a protocol that is simple and efficient to usc to connect a few pre- defined agents on a single machine or on several machines on the same local area network. We also need the protocol to he an appropriate one to scale up to a scenario in which we have a large number (i.e, hundreds or even thousands) of communicating agents scattered across the global internet and who are dynamically coming on and off line. Communication Modes. KQML will support several modes of communication among agents along several independent dimensions. Along one dimension, it supports interac- tions which differ in the number of agents involved — from a single agent to a single agent (i.c., point-to-point), as well as messages from one agent to a set of agents (i.c., multi- casting). Along another dimension, it permits one to specify the recipient agents cither explicitly (c.g., by internet address and port number), by symbolic address (¢.g., 10 “to the the TRANSCOMMapServer” or even by a declarative description form of broadcast (¢.g,. “10 any KIF-speaking agents interested in airport locations”). A final dimension involves syn- chonicity — it must support synchronous (blocking) as well as (non-blocking) asynchronous communication. Syntactic assumptions. Messages in the content, message and communication layers will be represented as Lisp s-expressions. They will be transmitted between processes in the form of ascii streams. The forms at the content layer will depend on the content-language being used and may be represented as strings, if necessary. ‘The forms at the message and communication layer will he ascii representations of lists with symbopl as the first clement and whose remaining clements use the Common Lisp keyword argument convention. Security. Security is an issue in any distributed environment. We will need to develop conventions and procedures for authentication which will allow an agent to verify that another agent is who it purports to he. Comment: We should take advantage of the kerberos system for secure authctication being developed as a part of Project Athena. It appears to have all the right hooks and will be widely available on IBM, DEC, SUN and other Unix platforms. Transaction. Interactions among, knowledge-based systems have a different kind of trans- action processing which will require something other than the now standard two-phase commit. ‘That is because interacting agents may use information and knowledge gained from one information source for longer periods of time than read/write locks support. In ‘Scanned with CamScannerean eae an apa ore wr (eg. LOOM, Prolog) (eg. LOOM, Profog) coat | rmpercans | | Ae Aap ‘rept nnd srt tn wre me Tate terme Protoot || | rr [atta] Internet Protest Proce! ct et sine sim Figure 4: Modern internet communication is governed by a" protocol stack” with distinct, well-defined layers. Communication between intelligent agents should also be governed by a protocol stack with distinct, well-defined layers one way, knowledge-based systems are similar to other advanced systems such as software engineering or CAD/CAM design environments (see Computing Surveys, 1991). Further, interactions among knowledge-based systems may better he cast in terms of helicf spaces and/or logics of belief than in terms of low level transactions. The development of a good model to support transactions among intelligent agents is a research topic for the KQML group to consider sometime in the future, Developing a workable solution which is incremen- tally implementable may prove key to the ultimate success of the KQML. effort. Comment: Transactions for interacting knowledge-based systems are different than what is standardly thought of for conventional databases, Bul what is the relationship to versioned databases and more advanced applications like software enginecring and CAD/CAM databases. Protocol Approach. ‘The Knowledge Query and Manipulation Language (KQML) is a language and a protocol to support the high level communication among intelligent agen It can be used as a language for an application program to interact with an intelligent system or for two or more intelligent systems to interact cooperatively in problem solving. We argue that KQML should he defined as more than a language with a syntax and semantics, but must also include a protocol which governs the use of the language (e.g., a pragmatic component). Using a protocol approach is standard in modern communication and distributed pro- cessing. The first diagram in Figure 4 shows a simplified version of the standard protocol stack for network communication over an internet. At the top of the stack is the application- level protocol, in this case SMTP (Simple Mail Transfer Protocol) and at the bottom is the low level protocol in which data is actually exchanged, From a mailer’s point of view, it is communicating with another mailer using the SMTP protocol. It need not know any of the details of the protocols which support its communication. We are developing a similar approach to support communication among intelligent agents — defining a protocol stack for transferring knowledge across the internet. ‘The second diagram in Figure 4 shows a simple protocol stack we are using for the model of KQMIL. We assume that the KQMI protocol stack is an application protocol layer of the standard OSI model and thus assume reliable communication. SKTP, a Simple Knowledge Transfer Protocol, supports KQMI. interactions and is de- fined as a protocol stack with at least three layers: content, message and communication. ‘Scanned with CamScannerAdditional layers will appear below these three to supply reliable communication streams between the processes. The content layer contains an expression in some language which encodes the knowledge to be conveyed. The message layer adds additional attributes which describe attributes of the content layer such as the language it is expressed in, the ontology it assumes and the kind of speech act it represents (e.g... an assertion or a query). The final communication layer adds still more attributes which describe the lower level communica- tion parameters, such as the identity of the sender and recipient and whether or not the communication is meant to by synchronous or asynchronous. ‘Scanned with CamScannerKnowledge Interchange Format (KIF) Knowledge Interchange Format (KIF) is a computer- oriented language for the interchange of knowledge among disparate programs. It has declarative semantics (i.e. the meaning of expressions in the representation can be understood without appeal to an interpreter for manipulating those expressions); it is logically comprehensive (i.e. it provides for the expression of arbitrary sentences in the first-order predicate calculus); it provides for the representation of knowledge about the representation of knowledge; it provides for the representation of nonmonotonic reasoning rules; and it provides for the definition of objects, functions, and relations. ‘Scanned with CamScannerArtificial Intelligence and Negotiation? ¢ What is artificial intetligence? Thinking Rationally Acting Humanly Acting Rationally ‘Scanned with CamScanner* Synergies between both fields: Thinking Humanly Thinking Rationally INEGotiation) Sify Sai Systems PoJewoiny Pela (oho y tice a) (o) Cy ay ) 3 = 5 ter © 5} Acting humanly Acting Rationally ‘Scanned with CamScannerE-marketplaces: Automated Negotiation — Thinking Rationally & Acting Rationally — Goal: Optimality according to the available information * Pareto Optimality * Nash Bargaining Point — Examples: Ebay, Amazon, etc... — Approaches: + Algorithmic Game Theory * Bounded Rationality Approaches & Heuristics * Mechanism Design ‘Scanned with CamScanner¢ Social Simulation — Thinking Humanly & Acting Humanly — Goal: Mimick human behavior to provide predictions * Emotions, cultural factors, social identity theory, etc. — Examples: Supply chain simulation — Uses: « Pilot experiments * Train real negotiators « Predict the effect of new environmental conditions ‘Scanned with CamScanner* Negotiation Support Systems = All-rounder! — Goal: Support one/all parties to reach an efficient agreement = Examples: AutoMed, Persuader — Approaches: * Enforcement/Recommendation * Best Response Mechanism * Reasoning about the opponent ‘Scanned with CamScanner3.1, Integrative Bargaining Model The goal of integrative negotiation (also known as “non-zero-sum-game” or “win-win game”) is to attain a mutually-beneficial agreement that maximizes settlement efficiency and fairness under suitable circumstances [5]. Integrative approaches employ objective criteria to make a circumstance of mutual gain and emphasizes the significance of exchanging information among the negotiators [26]. Walton and McKersie (1965) described the integrative bargaining model as a negotiation approach in which negotiators employ problem-solving behavior that refers to a state of desire for finding a solution to the problem to reach a definite goal. Problem-solving is generally recommended to achieve an integrative settlement [27,28]. Negotiators attempt to redefine the problem, analyze the cause of settlement difficulties and explore a wide range of mutually-acceptable alternative solutions through maximum information sharing and disclosure of each party’s needs and interests [28-30]. The effectiveness of the problem-solving approach depends upon the presence of some psychological and information conditions: motivation, information and language, trust and a supportive climate [12]. Motivation describes how parties must have the motivation to solve the problem and thus anticipate the problem as significant enough to address and discuss [12]. Information and language state that those participating in the problem-solving process must have contact to appropriate information relevant and be authorized to use it [12]. Meanwhile, the trust and support climate is marked by encouragement and freedom to behave spontaneously without fear of sanctions [12]. ‘Scanned with CamScannerTA wa | cy \ Tat 7 m2 73 GI We change Figure 1. Research model 1. Sustainability 2019, 11, 6832 ‘situation Process Has Negotiation ‘Situation Outcome cy Counterpan's ‘Soci psychological we Outcome ‘Opponent Desire for Future Negotiation Figure 2. Research model 2. ‘Scanned with CamScanner