Artificial Intelligence and Agents
In the field of artificial intelligence, agents are also known as “Autonomous Agents”. They are adaptive systems capable of completing target tasks. Since the establishment of the artificial intelligence discipline in the 20th century, the design and implementation of agents have been a core goal for researchers in the field. The key feature of agents is their autonomy: agents can control their own behavior and internal states, make decisions, and execute corresponding actions based on changes in the external environment and target tasks, all without direct human intervention. Furthermore, agents need to effectively communicate and interact with humans to better serve them.
1. Agents Based on Traditional Artificial Intelligence
Agents in the field of artificial intelligence first need to have the ability to perceive the external environment to collect necessary information and understand the basic conditions and dynamic changes of that environment. Based on the collected external environmental information, agents can solve specific problems in target tasks through deterministic or non-deterministic logical reasoning. Additionally, agents can utilize the data collected from the external environment to automatically summarize objective patterns in the data through machine learning, thereby improving their effectiveness and efficiency in solving target tasks. In this process, agents can establish memory and retrieval mechanisms to support the storage and retrieval of data and past experiences, further enhancing their performance. On this basis, agents can weigh pros and cons using specific mathematical models and algorithms, make reasonable decisions regarding target tasks, and translate those decisions into actual actions and impacts on the external environment through execution mechanisms, ultimately completing the target tasks.
The design of agents is typically based on various traditional artificial intelligence technologies, including but not limited to speech recognition, computer vision, knowledge representation and reasoning, machine learning, natural language processing, and robotics. The specific implementation of agents can be in software form, such as mobile personal assistants, e-commerce recommendation systems, and online customer service robots; or in hardware form, such as autonomous vehicles, industrial robots, and service robots.
2. Agents Based on Multimodal Large Models
With the rapid development of generative artificial intelligence technology, especially the emergence of multimodal large models (hereinafter referred to as “large models”), new opportunities have arisen for the design and construction of agents. Large models refer to artificial intelligence models capable of processing and analyzing multimodal data inputs such as text, images, audio, and video. Large models, represented by GPT-4, typically have hundreds of billions of parameters and demonstrate exceptional performance in natural language processing and audio-visual analysis across multiple tasks. In recent years, researchers have begun to attempt to construct agents based on large models as core foundational support. Compared to traditional agents, agents based on large models possess significant advantages and characteristics, particularly in the following core capabilities.
(1) Multimodal Perception Capability.Large models can perceive various modalities of data such as images, speech, and text, achieving individual or integrated perception functions. Therefore, agents constructed based on large models demonstrate more intelligent multimodal perception capabilities, allowing them to comprehend their external environment more comprehensively. These agents can not only perform visual and auditory perception but also integrate information from different modalities to achieve a perception capability closer to human levels.
(2) Reasoning and Planning Capability.Large models possess strong logical reasoning abilities, allowing designers to stimulate the model to engage in deep thinking by setting “prompt information” to form coherent “thought chains” or systematic “thinking trees”. Thus, agents based on large models can decompose complex tasks into a series of actionable sub-tasks. Through multi-step logical reasoning, they can autonomously plan a series of actions to achieve target tasks, efficiently exploring and implementing strategies to solve complex problems.
(3) Learning and Decision-Making Capability.Large models, due to their massive parameter scale and complex artificial neural network architecture, accumulate a comprehensive understanding of the objective world through training on vast amounts of multimodal data, providing a solid knowledge base for effective decision-making. Furthermore, large models utilize techniques such as “fine-tuning” and “retrieval-augmented generation” to deeply learn specific features and knowledge for particular domains or tasks, reducing the occurrence of “hallucinations” (i.e., erroneous or inaccurate information). Therefore, agents based on large models can provide more comprehensive and in-depth information, ensuring the quality and reliability of the learning and decision-making processes.
(4) Multirole Interaction Capability.Large models have the ability to understand and capture contextual information in multi-turn dialogues, allowing them to analyze dialogue situations and user intentions, generating logically coherent and appropriate responses accordingly. Therefore, agents constructed based on large models can interact more effectively with human users, external environments, and other agents. In human-computer interaction, agents can collaborate with human users while providing a high-quality user experience. In interactions with other agents, different agents can engage in effective discussions and collaborations based on their respective roles and functions, jointly promoting the efficient completion of complex tasks.
(5) Memory and Evolution Capability.Large models can utilize external memory and retrieval mechanisms of artificial intelligence systems to effectively store and retrieve specialized knowledge. As a result, agents based on large models can achieve self-reflection and self-improvement similar to human autonomous thinking by reflecting on and summarizing external memory and specialized knowledge. This self-evolution capability not only enhances the adaptability and flexibility of agents but also provides the possibility for continuous optimization, maintaining their advantages and efficiency in the ever-changing external environment and during the completion of complex tasks.
Agents for the Education Sector
The deep integration of artificial intelligence and education has given rise to agents specifically designed for the education sector, known as educational agents. Due to the diversity of educational scenarios and the complexity of service targets, the construction of educational agents must meet the unique needs of the education sector and exhibit specialized characteristics and functions distinct from agents in other vertical fields.
1. Educational Task Setting
Educational agents need to set educational scenarios, educational needs, and educational roles based on target educational tasks. The setting of educational scenarios provides background and environmental information for educational tasks, such as project-based learning scenarios centered around students or traditional classroom teaching scenarios; the setting of educational needs provides specific goals and descriptions for educational tasks, such as driving questions for project-based learning or conducting evaluations of teachers’ classroom teaching abilities; and the setting of educational roles assigns specific role information that the educational agent needs to play in educational tasks, such as peer students or research experts.
2. Educational Task Planning
Guided by established educational tasks, educational agents need to autonomously plan tasks. First, educational agents can fully utilize the reasoning and planning capabilities of large models to autonomously conceive and design preliminary solutions based on the scenario and need information set in the educational task. On this basis, educational agents will decompose the generated overall plan into multiple actionable sub-tasks. Educational agents need to continuously monitor the actual implementation effects of each sub-task and receive feedback from educational users. If the target or expected effect is not achieved, educational agents need to dynamically adjust the solutions and sub-tasks to ensure the accomplishment of educational tasks.
3. Educational Task Implementation
For each planned sub-task, educational agents can effectively solve them based on the learning and decision-making capabilities of large models. For sub-tasks that exceed the direct handling capacity of the agent, educational agents need to actively invoke external third-party tools or query knowledge bases online for resolution. For example, for subjects like mathematics that require precise calculations, educational agents can actively connect to third-party calculation tools to ensure the accuracy and efficiency of calculations. To provide the latest educational resources and data, educational agents can also access professional educational resource public service platforms online to obtain the required information.
4. Educational User Interaction and Self-Evolution
During the process of completing target tasks, educational agents can interact with human users, other agents, and the educational environment, continuously self-evolving in the process. In interactions with humans, agents can gain insights into the needs of different educational users and roles, providing personalized and multimodal human-computer interaction experiences; in collaboration with other agents, they can jointly promote the completion of educational tasks through discussions and debates; in interactions with the educational environment, educational agents can perceive and respond to environmental changes, optimizing teaching setups. Moreover, after various interactions, educational agents can engage in self-reflection and continuous improvement through their memory and evolution mechanisms.
5. Educational Knowledge Base Construction
The achievement of educational tasks often requires high accuracy and interpretability to enhance educational users’ trust. Therefore, educational agents often rely on specialized educational knowledge and data when planning and executing educational tasks. To meet these requirements, a local specialized educational knowledge base needs to be constructed for educational agents, covering personalized information of users. Educational agents can utilize techniques like “retrieval-augmented generation” to efficiently leverage information resources from the local educational knowledge base, integrating them into generated answers or solutions, thereby improving the accuracy and reliability of educational information output. Additionally, through continuous interaction with the local knowledge base, educational agents can continuously learn and adapt, expanding their knowledge boundaries and enhancing their ability to handle various educational tasks. This ongoing learning and adaptation process helps agents better understand educational content and provide more personalized and targeted educational services.
Applications of Educational Agents
The research and development of educational agents are still in their early stages, but they have already begun to demonstrate significant application potential and important roles in typical educational scenarios such as classroom teaching, educational evaluation, and teacher research.
In classroom teaching scenarios, teachers and agents can form collaborative partnerships to jointly promote the implementation of teaching activities. Agents in the classroom assume diverse roles and functions, providing multidimensional support for teachers. For instance, in project-based learning teaching models, learners rely on the continuous support of teachers and peers to achieve project goals. During this process, educational agents can act as “teaching assistants” and “peers”, collaborating with teachers and learners to fully participate in various stages of project-based learning. This includes proposing personalized questions, collaboratively designing project plans, and co-creating project outcomes. Additionally, educational agents can automatically identify individual students’ emotional states during the collaboration process and provide corresponding emotional support and interaction feedback. The human-computer collaborative teaching supported by educational agents not only improves teaching efficiency but also promotes students’ proactive exploration and teamwork, playing a supportive role in future classroom teaching.
2. Educational Evaluation
In educational evaluation, educational agents can utilize their multimodal perception capabilities and memory functions to facilitate the shift of evaluation models from experience-dependent subjectivity to data-driven objectivity, thereby enhancing the objectivity and fairness of educational evaluations. For example, during the evaluation and presentation stages of project-based learning works, educational agents can utilize the process information stored in their memory modules regarding group collaboration to conduct detailed process evaluations of the project implementation. Simultaneously, educational agents can also implement teacher evaluations and peer assessments from the different perspectives of “teachers” and “peers”. Educational agents can pre-generate corresponding process and outcome evaluation rubrics based on personalized driving questions and project plans.
With the assistance of educational agents, teachers can collaboratively prepare lessons and engage in research activities. Educational agents can first provide teachers with teaching resources, including teaching plans and multimodal teaching materials, and simulate and demonstrate classroom scenarios, assisting teachers in iterating and optimizing lesson materials using external tools and resources to create teaching content that better meets students’ actual needs. After teachers utilize these teaching resources to complete teaching activities, educational agents can quantify and analyze the objective dimensions of teachers’ teaching abilities based on their perceptions and memories of classroom implementation, providing detailed analysis reports. In research activities, educational agents can also act as experts from various subject backgrounds, executing observation and evaluation tasks, conducting multidimensional analysis and evaluation of the teaching process. Teachers can interact and even debate with educational agents to improve their teaching methods and further optimize the teaching process.
The construction of agents has always been a long-term pursuit in the field of artificial intelligence. The development of educational agents requires a deep understanding of educational resources, learner characteristics, and teaching processes, supported by solid foundations in educational theory and learning science. Educational agents should possess the ability for continuous learning and self-improvement, achieving autonomous evolution and effective interaction among multiple agents through interaction with educational stakeholders. At the same time, the design and application of educational agents must comprehensively consider their profound impact on education and potential challenges. Designers should focus on enhancing the reliability and credibility of educational agents, ensuring they provide unbiased and fair educational services, while thoughtfully considering their role in shaping learners’ values and ethics.
Author Unit │Beijing Normal University, Faculty of Education, School of Educational Technology
Source │“Information Technology Education in Primary and Secondary Schools”, 2024, Issue 10
Click the image above to enter the online subscription page of the magazine
Recently, some teachers have received the following manuscript notification and certificate, both of which are counterfeit documents and seals.
Special Reminder! Teachers must not believe this! Do not transfer payments!
↓↓↓
For submissions, please recognize the official submission email ([email protected]) and the WeChat mini program “Information Technology Education Magazine for Primary and Secondary Schools“.
Typesetting | Yi Dan Preliminary Review | Zhang Ying
Re-examination | Lu Qiuhong Final Review | Su Jinzhuzhu
Like this article, don’t forget to click “See” to support~
New Issue Delivery│2024 Issue 10Magazine Directory
Benefits are here! Welcome to subscribe to “Information Technology Education for Primary and Secondary Schools” 2024 magazine
Call for Papers! Welcome to participate in the 2024 “Digital Empowerment of Primary and Secondary School Teaching Practice” paper collection activity
The 2024 Primary and Secondary School Students’ Information Technology Innovation Application Works Collection Activity is starting!
