2.2.1 BDI model [7].
The belief-desire-intention (BDI) model is a cognitive architecture used to model IA decision-making and reasoning processes. The BDI provides a framework for understanding how agents process information, set goals, make decisions, and take actions in complex and dynamic environments. It is widely used in the field of AI, in particular in the development of IA and MAS.
Beliefs reflect the IA's perception of the world. They encompass the agent's knowledge of the current state of the environment, including information about itself, other agents, objects, events, and so on. Beliefs are usually presented as a set of propositions or statements.
Desires represent the agent's goals or tasks. They indicate what the agent wants to achieve or do in the environment. Desires can range from short-term goals, such as performing a certain action, to long-term goals, such as maximizing utility or achieving a certain outcome.
Intentions encompass the planned or anticipated actions of an IA to achieve one's desires. An agent forms intentions by choosing a course of action that he believes will help achieve his goals. Intentions guide the agent's decision-making and behavior by establishing a specific plan of action.
Components of the BDI process.
- Belief revision. IAs constantly update and revise their beliefs based on new information and observations of the environment.
- Desire selection. IAs prioritize their desires based on their current beliefs and preferences. Desires can be conflicting, so agents must choose which one to follow.
- Intention formation. IAs make intention decisions by choosing a plan or course of action that they believe will lead to the achievement of their desires.
- Plan execution. IAs carry out their chosen plans to realize their intentions and achieve their goals.
- Plan monitoring and reconsideration. IAs control their environment and the execution of their plans. If unexpected events occur, agents can revise their intentions and form new plans accordingly.
BDI provides a structured approach to modeling IA behavior. It is suitable for scenarios where agents need to reason about complex and uncertain situations, make decisions based on their beliefs and desires, and adapt to changing environments.
BDI-based systems are commonly used in modeling MAS, robots, intelligent assistants, and decision support systems. The architecture allows the development of agents that exhibit goal-directed behavior, prioritize goals, and make context-appropriate decisions.
BDI draws inspiration from philosophy, in particular from the works of such philosophers as Michael Bratman and his theory of practical reasoning. Bratman's ideas about how people make decisions, form intentions and pursue goals influenced the development of the BDI architecture [7 p.15]. Early conceptualization of the BDI architecture can be traced back to the 1980s in the work of such researchers as Keith Decker, Anand S. Rao, and Michael Georgeff. They investigated ways to model IA behavior in MAS.
One of the pioneering contributions to the BDI architecture was the Practical Reasoning Systems (PRS) model developed by Michael Georgef and Amy Lansky in the late 1980s. PRS proposed the BDI framework to represent the practical reasoning and intention-based behavior of agents [7 p.22]. BDI gained popularity in the 1990s, when it was increasingly used in MAS development. It has been recognized as an effective approach for modeling the behavior of agents in complex environments where agents need to interact and cooperate.
Programming languages and frameworks have been developed to support the BDI architecture. One prominent example is the AgentSpeak family of languages, including AgentSpeak(L) and its variants. These languages provided a way to express the behavior and reasoning of agents based on BDI [7 p.235].
Over the years, the BDI architecture has been widely studied and applied in various fields, including multi-agent systems, robotics, intelligent learning systems, and decision support systems. This has contributed to advances in modeling complex behavior, collaboration, and decision-making. The BDI continues to evolve, taking into account research findings in fields such as CS, AI, and psychology. Extensions and variations of the BDI model are explored to find new applications.
2.2.2 ACT-R model [8].
ACT-R (Adaptive Control of Thought - Rational) is a cognitive architecture designed to model human cognition and behavior. Developed by John Anderson and his colleagues from Carnegie Mellon University. ACT-R offers a framework for understanding and modeling cognitive processes, including perception, memory, attention, learning, problem solving, and decision making. It aims to capture the underlying mechanisms of human cognition and behavior, making it a valuable tool for CS, AI research, and KBIA creation.
Concepts and features of ACT-R.
- Declarative and procedural memory. Divides memory into two components: declarative and procedural memory. Declarative memory stores facts, knowledge, and concepts, while procedural memory stores rules and processes for performing tasks.
- Production rules (production rules). Uses production rules to represent cognitive processes. These rules consist of conditions and actions that allow modeling of decision-making processes and agent behavior. ACT-R rules are related to procedural memory.
- Chunking and learning. Includes fragmentation, a mechanism for organizing and presenting information. Fragmentation allows IA to capture and reuse patterns of information, which contributes to more efficient cognitive processing, contains mechanisms for learning and obtaining new fragments.
- Buffers. Uses buffers as temporary areas of information storage during cognitive tasks. Buffers represent different aspects of cognition, such as sensory input, working memory, and goal information.
- Goal structure. Reflects goal-directed IA behavior. Goals represent the desired results of cognitive tasks that the agent seeks to achieve by performing the corresponding production rules.
ACT-R's focus on cognitive processes makes it a valuable tool for building intelligent systems that simulate cognitive processes, including perception, memory, attention, learning, and decision-making.
2.2.3 Soar model [9].
Soar is a cognitive architecture that serves as a computational model of human cognitive processes, reasoning, problem solving, and decision making. It was developed by John Laird, Allen Newell and Paul Rosenbloom from the University of Michigan. Over the years, it has been constantly improved and expanded. The purpose of the development was to cover the mechanisms underlying intelligent behavior, which makes it a valuable tool for modeling human cognition and creating intelligent agents based on knowledge [10].
Features and concepts of Soar.
Unified framework. It aims to provide a unified framework for modeling a wide range of cognitive processes, including perception, learning, memory, reasoning, and decision making.
Symbolic and subsymbolic processing. Uses both symbolic and subsymbolic processing to represent and manipulate knowledge: symbolic structures (such as productions) to represent high-level cognitive processes and subsymbolic connectionist mechanisms for learning.
Production systems It is based on the concept of production systems, which consist of production rules that represent models of reasoning, decision-making and behavior. Products have conditions (prerequisites) and actions (conclusions), which allows you to use them for modeling cognitive processes.
Working memory. Maintains the current state of the IA, including information about the environment, objectives, and related facts. The products are mapped to the contents of working memory to trigger actions.
Problem-space search. Uses search in the problem space to explore potential sequences of actions and infer the order of application of products that will achieve the desired goals.
Learning. Contains mechanisms for learning, including fragmentation (learning higher-level abstractions) and RL. Learning mechanisms allow the system to acquire new knowledge and improve its efficiency.
Subgoaling. Supports the creation of sub-goals and hierarchical plans, allowing the system to break down complex tasks into manageable sub-tasks.
Soar provides a comprehensive framework for modeling complex cognitive processes, making it suitable for building KBIAs that can exhibit human-like behavior. It bridges the gap between symbolic reasoning and subsymbolic learning.