The challenge of creating a real-life computational equivalent of the human mind is now attracting the attention of many scienti c groups from di erent areas of cybernetics and Arti cial Intelligence such as computational neuroscience, cognitive science, biologically inspired cognitive architectures etc. The paper presents a new cognitive architecture based on insertion modeling, one of the paradigms of a general theory of interaction, and a basis for multiagent system development. Insertion cognitive architecture is represented as a multilevel insertion machine which realizes itself as a high level insertion environment. It has a center to evaluate the success of its behavior which is a special type agent that can observe the interaction of a system with external environment. The main goal of a system is achieving maximum success repeated. As an agent this machine is inserted into its external environment and has the means to interact with it. The internal environment of intelligent cognitive agent creates and develops its own model and the model of external environment. If the external environment contains other agents, they can be modeled by internal environment which creates corresponding machines and interprets those machines using corresponding drivers, comparing the behaviors of models and external agents. Insertion architecture is now under development on the base of Insertion modeling system, developed in Glushkov Institute of Cybernetics.
General theory of interaction is a theory of information interaction in complex
distributed multi-agent systems. It has a long history. Contemporary part of
this history can be considered as starting from neuro networks of
McCullochPitts [
Insertion modeling is a trend that is developing over the last decade as an
approach to a general theory of interaction of agents and environments in
complex distributed multi-agent systems. The rst works in this direction have been
published in the middle of 90s [
In recent years, insertion modeling has been applied to the development
of systems for the veri cation of requirements and speci cations of distributed
interacting systems [2, 12, 19{21]. The system VRS, developed in order from
Motorola, has been successfully applied to verify the requirements and speci cations
in the eld of telecommunication systems, embedded systems, and real-time
systems. A new insertion modeling system IMS [
This paper presents the main principals of insertion modeling and the
conception of cognitive architecture based on insertion modeling. To understand the
formal part of the paper reader must be familiar with the concepts of labeled
transition system, bisimilarity and basic notions of general process theory. The
mathematical foundation of insertion modeling is presented in [
Insertion modeling deals with the construction of models and study the interaction of agents and environments in complex distributed multi-agent systems. Informally, the basic principles of the paradigm of insertion modeling can be formulated as follows. 1. The world is a hierarchy of environments and agents inserted into them. 2. Environments and agents are entities evolving in time. 3. Insertion of agent into environment changes the behavior of environment and produces new environment which is in general ready for the insertion of new agents. 4. Environments as agents can be inserted into higher level environment. 5. New agents can be inserted from external environment as well as from internal agents (environments). 6. Agents and environments can model another agents and environments on the di erent levels of abstraction.
All these principles can be formalized in terms of transition systems, behavior algebras, and insertion functions. This formalization can be used as high level abstractions of biological entities needed for computer modeling of human mind.
The rst and the second principals are commonly used in information
modelling of di erent kinds of systems, for example as in object oriented or agent
programming. They are also resembling to M. Minsky's approach of the society
of mind [
The third principal is clear intuitively, but has a special re nement in insertion modelling. We treat agents as transition systems with states considered up to bisimilarity (or up to behavior, which is the same). The type of an agent is the set of actions it can perform. The term action we use as a synonim of label for transitions, and it can denote signals or messages to send, events in which an agent can participate etc. This is the most general notion of agent which must be distinguished from more special notions of autonomous or intellectual agents in AI.
Transition system consists of states and transitions that connect states.
Transitions are labeled by actions (signals, events, instructions, statements etc.).
Transition systems are evolving in time changing their states, and actions are
observable symbolic structures used for communication. We use the well-known
notation s a! s0 to express the fact that transition system can evolve from the
state s to s0 performing action a. Usually transition systems are nondeterministic
and there can be several transitions coming from the same state even labeled by
the same action. If we abstract from the structure of states and concentrate only
on (branching) sequences of observable actions we obtain the state equivalence
called bisimilarity (originated from [
Environment by de nition is an agent that possesses the insertion function. Given the state of environment s and the state of agent u, insertion function computes the new state of environment which is denoted as s[u]. Note that we consider states up to bisimilarity and if we have some representation of behaviors, the behaviors of environment and agent can be used as states. The state s[u] is a state of environment and we can use insertion function to insert a new agent v into environment s[u] : (s[u])[v] = s[u; v]. Repeating this construction we can obtain the state of environment s[u1; u2; : : :] with several agents inserted into it. Insertion function can be considered as an operator over the states of environment, and if the states are identi ed with behaviors, then the insertion of a new agent changes the behavior of environment.
Environment is an agent with insertion function, so if we forget the insertion function, then environment can be inserted as agent into a higher level environment and we can obtain hierarchical structure like
s[s1[u11; u12; : : :]E1 ; s2[u21; u22; : : :]E2 ; : : :]E Here notation s[u1; u2; : : :]E explicitly shows the environment E to which the state s belongs (environment indexes can be omitted if they are known from the context). This re nes the fourth principle.
Environment is an agent which can be inserted into external environment and having agents inserted into this environment. The evolution of agents can be de ned by the rules for transitions. The rules s[u] a! s[u; v] and s[t[u; v]] a! s[t[u]; v] can be used for the illustration of the 5-th principal.
We consider the creating and manipulation of the models of external and internal environments as the main property of cognitive processes of intellectual agent. Formalization of this property in terms of insertion modeling supports the 6-th principal.
Cognitive architecture will be constructed as a multilevel insertion environment. Below we shall de ne the main kinds of blocks used for construction of cognitive architecture. They are local description unites and insertion machines. 3
To represent behaviors of transition systems we use behavior algebras (a kind of process algebra). Behavior algebra is de ned by the set of actions and the set of behaviors (processes). It has two operations and termination constants. Operations are pre xing a:u (a - action, u - behavior) and nondeterministic choice u + v (u and v - behaviors). Termination constants are successful termination , deadlock 0, and unde ned behavior ?. It has also approximation relation , which is a partial order with minimal element ?, and is used for constructing a complete algebra with xed point theorem. To de ne in nite behaviors we use equations in behavior algebra. These equations have the form of recursive de nitions ui = Fi(u1; u2; : : :); i = 1; 2; : : : and de ne left hand side functions as the components of a minimal xed point. Left hand sides of these de nitions can depend on parameters ui(x) = Fi(u; x) of di erent types. In complete behavior algebra each behavior has a representation (normal form) u =
X ai:ui + "i i2I which is de ned uniquely (up to commutativity and associativity of nondeterministic choice), if all ai:ui are di erent ("u is a termination constant).
The type of environment is de ned by two action sets: the set of environment actions and the set of agent actions. The last de nes the type of agents which can be inserted into this environment: if the set of agent actions is included in the set of agent actions of environment then this agent can be inserted into this environment. This relation is called compatibility relation between agents and environments (agent is compatible with environment if it can be inserted into this environment). Multilevel environment is a family of environments with distinguished the most external environment. The compatibility relation on the set of environments de nes a directed graph and we demand for multilevel environment that the outermost environment would be reachable from any environment of the family in this graph.
To de ne the insertion function for some environment it is su cient to de ne transition relation for all states of environment including states with inserted agents. The common approach is to de ne behavior by means of rules. The following is an example of such rule: s !b s0; u !a u0 s[u] !c s0[u0]
P (a; b; c) This rule can be interpreted as follows. Agent in the state u can make a transition u !a u0. Environment allows this transition if the predicate P (a; b; c) is true. This rule de nes behavior property of environment in some local neighborhood of the state s[u]. So such a rule belongs to the class of local description units discussed in the next section.
At a given moment of time an agent belongs (is inserted) to only one environment. But if the type of an agent is compatible with the type of another environment it can move to this environment. Such a movements can be described by the following types of rules: moving from internal to external environment; u moveup E
! u0 E[F [u; v]; w] moveup(F !E!) E[F [v]; u0; w]
u movedn !F u0 E[F [v]; u; w] movedn(E!F !) E[F [u0; v]; w] P1(E; F; u; moveup(E))
P2(E; F; u; movedn(F ))
moving from external environment to internal one;
E[F [u; v]; G[w]] moveto(F !G!) E[F [v]; G[u0; w]]
P3(E; F; u; moveto(F ))
moving to another environment on the same level. In all cases permitting
conditions must include the compatibility conditions for corresponding agents and
environments. The rules above de ne the property of a system called mobility
and underlies the calculus of mobile ambients of Luca Cardelli [
A special type of environments is considered in cognitive architecture to have a su ciently rich language for the description of environment states properties. These environments are called attribute environments. There are two kinds of attribute environments { concrete and symbolic.
The state of concrete attribute environment is the valuation of attributes symbols that change their values while changing the state in time. Each attribute has type (numeric, symbolic, enumerated, agent and behavior types, functional types etc.). Some of functional and predicate symbols are interpreted symbols. Now logic formulas can be used for the description of properties of agent or environment states. We use the rst order logic formulas as the basis that can be extended by fuzzy logic, temporal logic etc.
The general form of local description unit is the following:
8x( (x; r) !< P (x; r) > (x; r)); where x is a list of typed parameters, r is a list of attributes, (x; r) and (x; r) are logic formulas, < P (x; r) > is a process - nite behavior of an environment. Local descriptions can be considered as formulas of dynamic logic, or Hoare triples, or productions - the most popular units of procedural memory in AI. In any case they describe local dynamic properties of environment behavior: for all possible values of parameters, if precondition is true then a process of a local description unit can start and after successful termination of this process a postcondition must be true.
The states of symbolic environment are formulas of basic logic language of environment. Such formulas are abstractions of classes of concrete states. Each symbolic state covers the set of concrete states and the traces generated by local description units cover the sets of concrete traces.
Local description units are the main units of knowledge representation in cognitive architecture. A set of local description units can be used for the de nition of transitions of environment. In this case they can be considered as procedural knowledge units. Logic knowledge can be represented as environment with the states representing the current knowledge, and the local description units corresponding to the rules of inference in corresponding calculus. Local description units can be applied in forward and backward modes. Forward mode can be used for the generating of new knowledge, backward mode { for answering queries.
Another construction blocks for cognitive architecture are insertion machines intended for implementation of insertion environments. The input of insertion machine is the description of a multilevel environment (a model of an environment) and its initial state, an output depends on the goal that is put to machine.
Multilevel environments are represented in cognitive architecture by means of environment descriptions for di erent levels and a set of local description units for insertion functions. Environment description contains the signature of environment that includes types of attributes, types of inserted agents, and also the description of sets of environment and agent actions. Local description units used for the de nition of insertion function are organized as a knowledge base with special data structures providing e cient access to the needed descriptions and history of their use.
To implement multilevel environment di erent kinds of insertion machines are used. But all of them have the general architecture represented on the Fig.1. Three main components of insertion machine are model driver (MD), behavior unfolder (Unf), and interactor (Intr). Model driver is a component which controls the machine traversal along the behavior tree of a model. The state of a model is represented as a text in the input language of insertion machine and is considered as an algebraic expression. The input language includes the recursive de nitions of agent behaviors, the notation for insertion function, and possibly some compositions for environment states. Before computing insertion function the state of a system must be represented in the form s[u1; u2; : : :]. This functionality is performed by agent behavior unfolder. To make the movement, the state of environment must be reduced to the normal form
X ai:ui + " i2I where ai are actions, ui are environment states, " is a termination constant. This functionality is performed by the module environment interactor. It computes the insertion function calling recursively if it is necessary the agent behavior unfolder.
Two kinds of insertion machines are distinguished: real time or interactive and analytical insertion machines. The rst ones are functioning in the real or virtual environment, interacting with it in the real or virtual time. Analytical machines intended for model analysis, investigation of its properties, solving problems etc. The drivers for two kinds of machines correspondingly are also divided into interactive and analytical drivers. Interactive driver after normalizing the state of environment must select exactly one alternative and perform the action speci ed as a pre x of this alternative. Insertion machine with interactive driver operates as an agent inserted into external environment with insertion function de ning the laws of functioning of this environment. External environment, for example, can change a behavior pre x of insertion machine according to their insertion function. Cognitive interactive driver has criteria of successful functioning in external environment, it accumulates the information about its past in episodic memory, develops the models of external environment, uses some learning algorithms to improve the strategy of selecting actions and increase the level of successful functioning. In addition it should have specialized tools for exchange the signals with external environment (for example, perception of visual or acoustical information, space movement etc.).
Analytical insertion machine as opposed to interactive one can consider different variants of making decisions about performed actions, returning to choice points (as in logic programming) and consider di erent paths in the behavior tree of a model. The model of a system can include the model of external environment of this system, and the driver performance depends on the goals of insertion machine. In the general case analytical machine solves the problems by search of states, having the corresponding properties (goal states) or states in which given safety properties are violated. The external environment for insertion machine can be represented by a user who interacts with insertion machine, sets problems, and controls the activity of insertion machine. Analytical machine enriched by logic and deductive tools are used for generating traces of symbolic models of systems. The state of symbolic model is represented by means of properties of the values of attributes rather than their concrete values.
Insertion machine with separated external environment interface can be implemented as a transition system with hidden structure that separates the kernel environment state and the states of inserted agents. Such implementation can be more e cient and can be constructed using partial computations or other specialization and optimization programming tools. 6
Like well-known cognitive architectures such as Soar [
On the abstract level ICAR has the same architecture as cognitive agents that can be created in it. From this point of view it can be considered as an intellectual assistant for user who interacts with ICAR in the process of creating cognitive agents. The general architecture of cognitive agent of ICAR is represented on Fig.2.
In general cognitive agent is constructed as a real time multilevel insertion machine which realizes itself as a highest level internal environment. As an agent, this machine is inserted in its external environment and has the means to interact with it. This external environment includes a user and objects of external (real or virtual) world to which agent has access.
One or several self-models can be inserted into the internal environment of cognitive agent to be used when interacting with external environment or making decisions and planning future activities. An agent has an estimation mechanism to evaluate the success of its behavior. This mechanism is realized in the form of a special agent that can observe the interaction of a system with external environment and make estimation according to some criteria. These criteria can be prede ned initially and evolves in the future according to obtained experience. The main goal of a system is achieving maximum success repeated.
The self-models of cognitive agent are created and developed together with the models of external environment. If the external environment contains other agents, they can be modeled by internal environment which creates corresponding machines and interprets those machines using corresponding drivers, comparing the behaviors of models and external agents. All these models are evolving and developing in the process of accumulating the experience in interaction with the external world.
Some mechanisms that model emotional or psychological features (humor and concentration, pleasure and anger, etc.) can be implemented at higher levels of cognitive strucure. The mechanisms of decision making, planning and executing plans are also at higher levels.
The main part of cognitive structure is the base of models describing the history of cognitive agent functioning at di erent levels of abstraction. The interface with external world provides language (symbolic) communication and image processing. All interaction histories are processed in the working memory of the self-level insertion machines and then transferred to the appropriate levels of a model base.
The model base is always active. The analytical insertion machines which control and manage the structure of model base are always busy with searching solution of problems and performing tasks with ansatisfactory answers, or creating new models. All this activity models subconcious levels of cognition and time-to-time interact with the higher levels of cognitive structure. Independent levels of cognitive structure are working in parallel.
The hierarchy of environments of cognitive agent in some sense are similar to six layers of neocortex. Moving from low levels to higher ones the levels of abstraction are increased and used more and more abstract symbolic models. How to create such models is a big challenge and we are working on it now.
Cognitive analytical insertion machines of ICAR are used by cognitive agents to learn their models and their interaction with external environment to solve problems better, accepts user helps as a teacher and teach user how to interact better with ICAR. General learning mechanisms are the parts of model drivers of di erent types.
The description of cognitive architecture in the last section is a very tentative
re ection of our far goals. The nearer goals include the further development
of our system of proving pogram correctness [