Q. “What do you give a hurt lemon?”
A. “Lemon aid”
Like me, you may have thought that the writer of this joke is a student. Actually, the joke writer in this case is Artificial Intelligence software – a “joke generator” called JAPE.
Artificial Intelligence (AI) has growing implications for schooling, and this article aims to set out some of AI’s main concepts, and explore how they can be applied to improving learning.
What is Artificial Intelligence?
Artificial Intelligence is a mature field in Computer Science that has delivered many innovations, for example:
- Deep Blue, the chess program that beat Kasparov
- “iRobot Roomba” automated vacuum cleaner, and “PackBot” used in Afghanistan and Iraq wars
- Spam filters that use Machine Learning
- Question answering systems that automatically answer factoid questions
AI is best known for aiming to reproduce human intelligence. The field was founded on the claim that intelligence can be simulated by a machine. Essentially AI is the design of systems that perceive their environment and take action that maximize their chances of success. AI addresses natural language processing, reasoning, knowledge, planning, learning, communication, perception and the ability to move and manipulate objects. AI is about many things including interacting with the real world; reasoning and planning; learning and adaptation.
There are several approaches to AI including:
- Building models of human cognition using psychology and cognitive science
- The logical thought approach with emphasis on “correct” inference
- Building rational “agents” – a computing object that perceives and acts
Key areas of application of AI in education include:
- Expert systems
- Intelligent tutoring systems
- Search, question and answers
Key AI Concepts
An initial view of AI reveals a field that is deeply divided into seemingly unrelated subfields. Some of these sub-fields even appear contradictory. For example, Neural Network techniques are considered by some a better model of human reasoning than rule-based Expert Systems, so lets take a closer look at these two approaches.
This approach mimics the human brain through the use of “nodes”, which resemble neurons. Neural Network technology – which uses layers of “input”, “hidden (process)” and “output” nodes – has been applied successfully to speech recognition, image analysis, adaptive control, games and robots. Most of neural networks are based on statistical estimations, classification optimization and control theory. Neural networks can be programmed to model the behavior of natural systems – e.g. responding to stimuli.
Expert Systems emulate the decision-making ability of a human expert by reasoning about knowledge – as opposed to following the procedures set out by a software developer as is the case of conventional programming. An expert system is divided into three parts – a knowledge base; an inference engine; and a dialog interface to communicate with users.
Neural Networks can be applied to the problem of Machine Learning – the design and development of algorithms that allow computers to evolve behaviors based on data from sensors, input devices, or databases. An important task in Machine Learning is pattern recognition, in which machines “learn” to automatically recognize complex patterns, and to make intelligent predictions.
In games which have concrete rules and multiple permutations – eg Chess – Machine Learning calculates the most likely outcomes of the game given the position on the board by playing simulated games into the future. In addition, pattern recognition enables the game to analyze the relative merits of different moves in the game, based on which ‘shapes’ were created by experts in historical games.
An intelligent agent is a set of independent software tools linked with other applications and databases running within one or several computer environments. Agent based technology systems include a degree of autonomous problem-solving ability. The primary function of an intelligent agent is to help a user better use, manage, and interact with a system or application. Additionally, software agents, like human agents (for example, an administrative assistant), can be authorized to make decisions and perform certain tasks.
Coach Mike, is an Intelligent Agent used at the Boston Museum of Science. Coach Mike’s job is to help visitors at Robot Park, an interactive exhibit for computer programming. By tracking visitor interactions and through the use of animation, gestures, and synthesized speech, Coach Mike provides several forms of support that seek to improve the experiences of museum visitors. These include orientation tactics, exploration support, and problem solving guidance. Additional tactics use encouragement and humor to entice visitors to stay more deeply engaged. Preliminary analysis of interaction logs suggest that visitors can follow Coach Mike’s guidance and may be less prone to immediate disengagement.
Herbert A. Simon, an AI pioneer, said – “If we understand the human mind, we begin to understand what we can do with educational technology.”
Human learning and reasoning is founded on multiple knowledge representations with different kinds of structures, such as trees, chains, dominance hierarchies, neighborhood graphs, and directed networks. From MIT Open Courseware (Image by Prof. Joshua Tenenbaum.)
With systems that can both “learn” and provide “expertise”, the implications of AI for schooling are profound. Whilst AI has potential for solving problems like optimal resourcing and improving operational performance, the strongest area for the application of AI in schooling is to make learning more effective.
AI in schooling can be traced back to 1967 when Logo was created. Since the introduction of Logo and “floor-bots” such as Turtles, ever more sophisticated robots along with associated control technologies such as Lego Mindstorms – have been used in schools. Products such as Focus Educational’s “BeeBot” is a recent addition to systems applying some of the principles of AI in a schooling environment.
AI in schooling is evolving in several different ways:
Question and Answer Systems (QA)
By 2020, we’ll be creating enough data for a stack of DVDs containing it to reach the moon and back three times! Regrettably, the quality of answers does not necessarily improve in proportional to the amount of information available. The current generation of search engines are essentially information retrieval systems providing a list of “hits” from which the user has to deduce the closest match. One of the goals of AI, therefore, is to enable more natural questioning resulting in better answers and related information.
The first QA systems were developed in the 1960s as natural-language interfaces to expert systems. Current QA systems first typically classify questions and then apply Natural Language Processing. Natural language ‘annotations’ describe content associated with ‘information segments’. An information segment is retrieved when its annotation matches an input question. A generating module then produces sentences – ‘candidate answers’. Finally, ‘answer extraction’ processes determine if the candidate answer does indeed answer the question.
The implications for QA systems in schooling are enormous and raise significant questions about the role of teachers, learning content and assessment.
Learning With Expert Systems
Imagine students being given the task of recognizing patterns on science laboratory slides and making correct classifications. By combining expert and pedagogic models we are able to exploit AI to “mash” both domain specific and more general learning principles into a rich learning experience. When classifying the slides, students will be not just presented with a “right or wrong” response, but their behavior will be refined through “machine understanding” of why the student is making their decisions. AI differs from more conventional computing approaches by being able to generate and handle both “feed-forward” and “feed-back”.
Taking this a step further are Intelligent Tutors. These record their interactions with students to better understand how to teach them. Computer tutors are capable of recording both longitudinal data, as well as data at a fine-time scale, such as mouse clicks and response time data. Using these interactions as a source of data to be mined provides a new view into understanding student learning processes.
Games and Simulations
Currently, the area in which AI is applied the most is Computer Games – and by a large margin. The use of scenario-based simulations and serious games for training has been well-accepted in many domains. Simulations require active processing and provide intrinsic feedback in an environment in which it is safe to make mistakes. Artificial ecosystems – like the one shown below – have proved popular and have their uses in schooling.
An interesting learning mechanism used in game based learning that is potentially usable in other contexts is “Transfer Learning” – which can help improve the speed and quality of learning. The idea is to use knowledge from previous experiences to improve the process of solving a new problem.
Two key AI methods underpin this approach –
- Case-Based Reasoning (CBR) – a set of techniques for solving new problems from related solutions that were previously successful.
- Reinforcement Learning (RL) – set of algorithms for solving problems using positive or negative feedback from the environment.
Reinforcement Learning can be delivered through the following mechanism –
- A central database with a collection of rules, mapping all possible actions and relative values.
- A learning component that takes feedback from the environment, and updates the utility value of each action. This is done using a reinforcement learning policy which estimates if there were any improvements since the last step.
- A planner then takes these rules, and computes a plan of action randomly based on the utility of the actions.
To anyone who has explored managed learning, this should sound quite familiar.
Two interesting models for understanding human learning in AI and Games context have come out of Microsoft Research:
This model classifies different types of learning in the context of games environments, but has transferability to broader understandings of the interface between computing and learning:
This model helps visualize the relative ease with which a game player can learn, depending on the granularity of detail presented to them:
- Too coarse: cannot learn a good policy
- Too fine: impossible to learn from little experience
- Just right: learn a good policy from little experience
Ramona Pierson, Chief Scientific Officer for Promethean, talks about ‘mashable’ digital content with embedded assessments tightly coupled to the curriculum, and learning progressions made ‘dynamic’ by AI. This can adjust learning progressions continually for each student, presenting cross-curriculum content and learning strategies based on a dynamic learning process.
“Imagine how powerful it would be for a student to have a customised textbook, sequencing of lessons, and embedded assessments that dynamically changed to ensure that he/she masters the material in the way that makes sense, and would result in obtaining nationally set benchmarks and learning outcomes”. (Mass Customisation And Personalisation Of Learning, Education Technology Solutions).
Nick Fekos, a former AI programmer in the financial sector and now at Athens College, agrees and is formulating plans for an intelligent object oriented knowledgebase that ‘learns’ from ‘experience’ and adjusts accordingly. The system Nick has in mind will implement dynamic, self-organizing and differentiated learning paths. The more the learning algorithm is used, the better it will get – perhaps something that can be said for the more general application of AI to schooling itself.
So How Do I Build an AI System?
Firstly, there is plenty of opportunities for getting students developing AI systems.
Besides Logo, its worth looking into Kodu – a visual programming language made specifically to enable children to create games.
Also check out Microsoft Robotics Developer Studio which helps make it easy to develop robot applications. The current version (4), which is in Beta, provides extensive support for the Kinect sensor hardware allowing developers to create Kinect-enabled robots in both a ‘Visual Simulation Environment’ and real-life.Integrating AI into other learning workloads is an altogether more complex task.
For anyone wanting to understanding the mechanics of programming an AI system, this excellent article shows how to programme a neural network in C#.
For a more comprehensive desicription, including important architectural principles, check out this paper from University of Southern California which explains how to build a simulation to teach soft skills such as negotiation and cultural awareness.
For a comprehensive coverage of the field of AI in Education, look at the proceedings from Artificial Intelligence in Education, 15th International Conference, AIED 2011, Auckland, New Zealand, June 28 – July 2011.
For a comprehensive coverage of the field of Intelligent Tutoring, look at the proceedings from the 10th International Conference on Intelligent Tutoring Systems, ITS 2010, Pittsburgh, PA, USA, June 14-18, 2010