According to the viewpoint of World Model, we discuss the deficiency of A General Model of Learning Agents proposed by Russell. We have to deliberate the importance of self-learning ability in order to make agent get closer to human real thinking intelligence. After completely analyzing and collecting, we proposed a Cognition Learning Model which based on using the self-schema to be agent’s internal learning focus and which could be compatible with the former agent system. By the model, agent pays learning attention to not only the World Model but also Self-Schema.
Because of the achievement of inner learning by the awareness of self-schema, agents could be closer to human thinking intelligence. Also the model offers a brand-new designing concept to solve the high-level intelligent problem which the agent of traditional world model cannot solve.
3.1. The Proposed Model: ACLM
World Model Learning is a concept which put the learning focus (Attention) at the outer environment. Basically the outer world is the entire of learning. Actually it focuses on the association between outer incentives and behavioral responses but not discuss about the internal operations. After continuous adapting to environment and world, agent could acquire various skills and working strategies. Some common agent learning models (e.g. Reinforcement Learning 、 Neural Network 、 Evolutionary Computing、Classifier System、Decision Tree), for example, we could explore that all of them used World Model to be the learning focus (means entire) from these famous models of traditional artificial intelligence.
To discuss further into the example, Evolutionary Computing didn’t include the viewpoint of internal learning. All its learning focus was put on the process for environment interaction. Through the pressure of existence given by evolution, it imitated other agents persistently, and through the process of crossover, mutation, and selection, it adapted to the environment, that is to say, Learning equals Adaptation.
Although these World Model learning methods have good outcome, it’s still not enough to simulate such human thinking intelligence on these models. In addition, more and more scholars started to research agent’s emotion and mental state, agent’s internal learning has to be regarded that is to pay attention on the self learning methods.
As the result of the analysis to World Model, the current designing concept has to be modified. Based on Russell’s general agent model, we proposed a new agent cognition learning model (Figure 7). The model is composed of three elements (Performance element 、 World Model Learning element 、 Self-schema Cognition element). Performance element, such as Russell’s design, is responsible for selecting external actions. The World Model Learning element is in charge of integrating the traditional learning components which only focused on outer environment to improve the learning efficiency. It takes some knowledge about the learning element and some feedback on how the agent doing, and determines how the performance element should be modified to do better in the future. Finally Self-Schema Cognition element, which is according to the past experience to get information into a knowledge structure, could help agent to understand, explain, and predict self-behavior. Because of the model design, World Model Learning and Self-Schema Cognition can coordinate each other, and present the most favorable method to improve performance. Eventually agent will be in possession of both external and internal learning concepts.
Figure 7: Agent Cognition Learning Model
The Agent cognition Learning Model we proposed could offer a new agent designing framework. Agent will possess external learning and internal cognition concepts of both World Model and Self-Schema. Therefore, the agent could self-discovery from self-awareness by adding various schemas that designers needed, and could effectively improve and promote the efficiency through the co-ordination between external learning and internal cognition, and then get closer to human thinking intelligent model.
3.2. Independence and Compatibility
Now the independence and compatibility of Agent Cognition Learning Model will be discussed in this section. If the model we proposed cannot be compatible with the former agent system, that is to say, we have to change the former agent design model before operation, it is unreasonable and uneconomical. Even the model is a perfect and complete one still cannot be accepted by the public.
Basically we separated the World Model and Self-schema completely based on the viewpoint of the external learning (learning theory) and internal cognition (cognition theory) in ACLM. Russell’s model for example, we combined all World Model related components into the World Model Learning element, and embedded the self-schema cognition element which is independence of World Model Learning element in addition.
Therefore, designers who hope to implement our ACLM in their own models don’t have to re-modify the whole original learning structure but embed the part of self-schema cognition design. That proves, no matter what World Model learning approach the designers applied, the ACLM would be compatible with all these agent systems as well.
Through the internal cognition concept proposed by ACLM, agents will have the human self-awareness ability. Self-awareness can help agents inwardly self-understand and self-identification and outwardly learn various important skills or change the working strategies. In that event, during the research betweens the computer simulation using agent-skill and artificial society, agents will be able to present human high-level complex internal thinking behavior. Moreover, the result of simulation will be more realistic and persuasive. In next chapter, we will experiment with ACLM by using a general conflict problem in artificial societies, so that the experimental results can help us to verify ACLM, which can be implemented and is compatible with former agent system.