Total: 31
This paper describes a fast hand strength estimation mod-el for the game of Gin Rummy. The algorithm is computationally inexpensive, and it incorporates not only cards in the player’s hand but also cards known to be in the opponent’s hand, cards in the discard pile, and the current game stage. This algorithm is used in conjunction with counterfactual regret (CFR) minimization to develop a gin rummy bot. CFR strategies were developed for the knocking strategies. The hand strength estimation algorithm was used to select a discard that balances the goals of maximizing the utility of the player’s hand and minimizing the likelihood that a card will be useful to the opponent. A study of the parameterization of this estimation algorithm demonstrates the soundness of approach as well as good performance under a wide range of parameter values.
Applications of Generative Machine Learning techniques such as Generative Adversarial Networks (GANs) are used to generate new instances of images, music, text, and videos. While GANs have now become commonplace on social media, a part of children’s lives, and have considerable ethical implications, existing K-12 AI education curricula do not include generative AI. We present a new module, “What are GANs?”, that teaches middle school students how GANs work and how they can create media using GANs. We developed an online, team-based game to simulate how GANs work. Students also interacted with up to four web tools that apply GANs to generate media. This module was piloted with 72 middle school students in a series of online workshops. We provide insight into student usage, understanding, and attitudes towards this lesson. Finally, we give suggestions for integrating this lesson into AI education curricula.
We report on the use of open-source natural language processing capabilities in a web-based interface to allow undergraduate students to apply what they have learned about formal natural language structures. The learning activities encourage students to interpret data in new ways, think originally about natural language, and critique the back-end NLP models and algorithms visualized on the user front end. This work is of relevance to AI resources developed for education by focusing on inclusivity of students from many disciplinary backgrounds. Specifically, we comprehensively extended a web-based system with new resources. To test the students' reactions to NLP analyses that offer insights into both the strengths and limitations of AI systems, we incorporated a range of automated analyses focused on language-independent processing or meaning representations which still represent challenges for NLP. We conducted a survey-based evaluation with students in open-ended case-based assignments in undergraduate coursework. Responses indicated that the students reinforced their knowledge, applied critical thinking about language and NLP applications, and used the application not to solve the assignment for them, but as a tool in their own effort to address the task. We further discuss how using interpretable visualizations of system decisions is an opportunity to learn about ethical issues in NLP, and how making AI systems interpretable may broaden multidisciplinary interest in AI in early educational experiences.
Games are an excellent tool for undergraduate research in artificial intelligence because they typically have clear objectives, a limited action space, and well-defined constraints. Nonetheless, games involving chance and imperfect information offer unique challenges for optimizing gameplay. In this paper, we analyze one such card game, gin rummy, and propose an artificial intelligence player based on empirically driven strategies. Our approach separates gameplay into three disjoint policies for drawing, discarding, and knocking. On each turn, decisions are influenced by offensive considerations as well as defensive moves. Tournament-style simulations enable us to determine statistically which combination of policies achieves the highest win rate. Our resulting player, dubbed Heisenbot, is competitive against strong baseline strategies.
In this article, we describe various approaches to opponent hand estimation in the card game Gin Rummy. We use an application of Bayes' rule, as well as both simple and convolutional neural networks, to recognize patterns in simulated game play and predict the opponent's hand. We also present a new minimal-sized construction for using arrays to pre-populate hand representation images. Finally, we define various metrics for evaluating estimations, and evaluate the strengths of our different estimations at different stages of the game.
Due to the computation time and resources required, there is no known optimal strategy for the game of Gin Rummy. Previous work in extensive games, such as Texas Hold ’em Poker, has found that hand fitness and information sets about the state of the game can be used to determine an improved strategy. These information sets, combined with algorithms for Counterfactual Regret Minimization, can arrive at a Nash Equilibrium strategy for smaller abstractions of extensive games. This paper builds on previous research by extending the premise of hand fitness to card fitness in the discard decision point of Gin Rummy. We argue that a card can be ranked based on whether it meets four specific characteristics at that stage in the game. These characteristics include its effect on deadwood points after one more turn, its utility to the opponent, and if it can contribute to a meld. An optimal discard choice can then be picked from the highest-ranked card by using a simplified Counterfactual regret minimization strategy that can be trained in less time due to its limited information set. While this does not look at every potential characteristic of card fitness, it outperformed other bots when evaluated in a large number of games. These bots did not consider card fitness as a whole, but rather considered characteristics separately. We argue that the characteristics defined are a part of the total information set that can determine the discard fitness of a card within a hand in the game of Gin Rummy.
Gin Rummy is a popular two-player card game involving choices to draw and discard cards to form sets of matching cards. Unlike other popular games such as Chess, Poker, and Go, there is little formal artificial intelligence research about how to make good decisions when playing Gin Rummy. In this paper, we develop an agent that plays Gin Rummy through a combination of known and expected card values, modeling the opponent to predict their cards of interest, and a conservative approach to assessing when to end the hand. In addition to discussing our observations about Gin Rummy that inspired our agent's design and how the agent works, we evaluate the relative importance of various features employed by our agent by competing agents which implement various subsets of those features.
Various Gin Rummy strategy guides provide heuristics for human players to improve their gameplay. Often these heuristics are either conflicting or contain ambiguity that limits their applicability, especially for discard and end-of-game decisions. This paper describes an approach to analyzing the machine learning capabilities of a Gin Rummy agent to help resolve these conflicts and ambiguities. There are three main decision points in the game: when to draw from the discard pile, which card to discard from the player's hand, and when to knock. The agent us-es a learning approach to estimate the expected utility for discards. An analysis of these utility values provides in-sight into resolving ambiguities in tips for discard decisions in human play. The agent’s end-of-game, or knocking, strategy was derived using Monte Carlo Counterfactual regret minimization (MCCFR). This approach was applied to estimate Nash equilibrium knocking strategies under different rules of the game. The analysis suggests that conflicts in the end-of-game playing tips are due in part to different rules used in common Gin Rummy variants.
Artificial Intelligence’s impact on society is increasingly pervasive. While innovative educational programs are being developed, there has been little understanding of how students, especially pre-college aged students, construct understanding and gain practice with core ideas about AI or what concepts are most appropriate for what age-levels. In this paper, we discuss a cognitive interview study with high school students to better understand how students learn AI concepts. We aim to shed light on questions including: what is the range of background knowledge and experiences students are able to apply in encountering AI concepts; what concepts are most readily accessible and which are more challenging; what misconceptions do students bring to bear on AI problems; and how to help students approach AI concepts by leveraging related concepts, such as mathematical and computational thinking). Results from the exploratory study have the potential to provide important insights into AI learning for pre-college youth. These initial findings can inform further investigations to ground the design of learning and assessment in evidence-based learning progressions and grade-level performance expectations.
We explore creating automated, personalized feedback in an intelligent tutoring system (ITS). Our goal is to pinpoint correct and incorrect concepts in student answers in order to achieve better student learning gains. Although automatic methods for providing personalized feedback exist, they do not explicitly inform students about which concepts in their answers are correct or incorrect. Our approach involves decomposing students answers using neural discourse segmentation and classification techniques. This decomposition yields a relational graph over all discourse units covered by the reference solutions and student answers. We use this inferred relational graph structure and a neural classifier to match student answers with reference solutions and generate personalized feedback. Although the process is completely automated and data-driven, the personalized feedback generated is highly contextual, domain-aware and effectively targets each student's misconceptions and knowledge gaps. We test our method in a dialogue-based ITS and demonstrate that our approach results in high-quality feedback and significantly improved student learning gains.
We demonstrate an AI agent for the card game of Gin Rummy. The agent uses simple heuristics in conjunction with a model that predicts the probability of each card's being in the opponent's hand. To estimate the probabilities for cards' being in the opponent's hand, we generate a dataset of Gin Rummy games using self-play, and train a random forest on the game information states. We explore the random forest classifier we trained and study the correspondence between its outputs and intuitively correct outputs. Our agent wins 61% of games against a baseline heuristic agent that does not use opponent hand estimation.
Body-tracking artificial intelligence (AI) systems like Kinect games, Snapchat Augmented Reality (AR) Lenses, and Instagram AR Filters are some of the most engaging ways students experience AI in their everyday lives. Additionally, many students have existing interests in physical hobbies like sports and dance. In this paper, we present PoseBlocks; a suite of block-based programming tools which enable students to build compelling body-interactive AI projects in any web browser, integrating camera/microphone inputs and body-sensing user interactions. To accomplish this, we provide a custom block-based programming environment building on the open source Scratch project, introducing new AI-model-powered blocks supporting body, hand, and face tracking, emotion recognition, and the ability to integrate custom image/pose/audio models from the online transfer learning tool Teachable Machine. We introduce editor functionality such as a project video recorder, pre-computed video loops, and integration with curriculum materials. We discuss deploying this toolkit with an accompanying curriculum in a series of synchronous online pilots with 46 students, aged 9-14. In analyzing class projects and discussions, we find that students learned to design, train, and integrate machine learning models in projects of their own devising while exploring ethical considerations such as stakeholder values and algorithmic bias in their interactive AI systems.
The constraint in sharing the same physical learning environment with students in distance learning poses difficulties to teachers. A significant teacher-student interaction without observing students' academic status is undesirable in the constructivist view on education. To remedy teachers' hardships in estimating students' knowledge state, we propose a Student Knowledge Prediction Framework that models and explains student's knowledge state for teachers. The knowledge state of a student is modeled to predict the future mastery level on a knowledge concept. The proposed framework is integrated into an e-learning application as a measure of automated feedback. We verified the applicability of the assessment framework through an expert survey. We anticipate that the proposed framework will achieve active teacher-student interaction by informing student knowledge state to teachers in distance learning.
With the rapid technological change of society with Artificial Intelligence, elementary schools' goal should be to prepare the next generations according to competencies. We propose an AI curriculum to cultivate students' AI literacy to answer the question of ‘why and what to teach’ on AI. The proposed AI curriculum focuses on achieving AI literacy based on three competencies: AI Knowledge, AI Skill, and AI Attitude. We anticipate that the proposed curriculum will equip students with core competencies for the future with AI.
We developed a heuristic-based reflex agent, Tonic, for the EAAI 2021 Undergraduate Research Challenge, which tasks competitors to create an autonomous player to play the card game gin rummy. Tonic's heuristics originate in expert knowledge and inform decision making for the three actions comprising a turn: drawing a card, discarding a card, and deciding when to knock. However, because these strategies are based in human intuition, there is often a lack of specificity to directly model them as algorithms. We developed parameterized models describing that intuition based on factors such as the number of turns played and an estimation of the opponent hand. To hone their performance, we conducted both manual analysis and parameter optimization (grid search) using self-play and play against a simple baseline agent. These heuristic models enable Tonic to win against the baseline agent at least 68% of the time.
The Contour to Classification game is a browser-based game that teaches middle school students basic concepts in supervised learning. The game is an online variant of the Neural Network game that was presented at AAAI Fall Symposium Teaching AI in K-12 track in 2019. We share preliminary findings from implementing the online version of the original Neural Network game in a pilot research study and describe the game’s evolution to the Contour to Classification game. The new game uses a simulation of a neural network to engage students, through digital drawing and selection interactions, in the classification of images. The players act as nodes in a multi-step process of compositing salient smaller features to form larger features and ultimately a partial contour of an object that is used to make a prediction. After evaluating the prediction, information is sent back through the network in processes mimicking back propagation and gradient descent. Additional rounds of the game can be played to witness how the network evolves and gets “better” at classifying images from contours. Through this game, we aimed for students to learn the structure, components, and functioning of a neural network, and the processes involved in supervised learning. The Contour to Classification game supports online student learning by providing the image classification experience using purely visual inputs to each layer. We will conclude with a discussion of if and how the evolving design addresses classroom needs and scaling considerations.
Artificial intelligence has emerged as a technology that is profoundly reshaping society and enabling rapid improvements in science, engineering, and mathematics, as well as information technology itself. This has generated increased demand for fostering an AI-literate populace as well as a growing recognition of the importance of promoting K-12 students’ awareness and interest in AI. Although efforts are be-ginning to incorporate AI learning within K-12 education, there is little research exploring how to introduce students to AI and how to support teachers to integrate AI learning experiences in their classrooms. This is especially true at the elementary school level. A particularly promising approach for providing effective and engaging AI learning experiences for elementary students is game-based learning. In this paper, we explore how to introduce AI-infused collaborative inquiry learning into upper elementary school (student ages 8 to 11) using game-based learning. To ground the work in the realities of elementary school classrooms, we present insights from interviews with elementary school teachers to under-stand how best to support them in integrating AI into their classrooms. We then present the design of PrimaryAI, a game-based learning environment that supports rich problem-based learning activities within upper elementary classrooms centered on AI applied toward solving life-science problems. Finally, we discuss some of the challenges we face in bringing AI-infused collaborative inquiry learning to upper elementary students.
We perform an empirical study of Gin Rummy knocking strategies, drawing insight from a population of AI players that vary in both discarding and knocking strategies. For our best performing player, simple linear regression yielded a knocking strategy that both affirmed the features expert players give attention to in making knock decisions, and yet called into question the way such features are conventionally used.
A vital part of any good strategy for most imperfect-information games is making predictions about the information that is unavailable. For example, in card games like Poker and Gin Rummy, predicting the kinds of cards the opponent is holding is necessary for playing well. Specifically, it is useful for agents to be able to map the partial game states that are made available to them to the probabilities of each of the possible complete game states, given that they are playing against other rational player(s). Finding this relationship, however, is difficult, as it requires knowledge of how a rational player would play, which is the problem this relationship is being used to solve. In this paper, we attempt to find this relationship in the context of the card game Gin Rummy, though instead of predicting the complete game state, we focus on what is most useful to a player: the opponent's hand. We do this by using heuristic utility functions to create an agent that approximates how a rational player would play, and then using the resulting game data to train a Deep Neural Network mapping known information to predictions about the opponent's hand. This model is used to improve the existing agent and, in turn, to produce more data to create better models.
This paper describes the design and training of a computer Gin Rummy player. The system includes three main components to make decisions about drawing cards, discarding, and ending the game, with numerous parameters controlling behavior. In particular, an ensemble approach is explored in the discard decision. Finally, three sets of parameter tuning and performance experiments are analyzed.
This paper describes a deterministic approach to building a fixed-strategy gin rummy player. In the paper, we develop and evaluate both heuristic and neural network models for informing draw, discard, and knock decisions in the game. In this empirical study, we test performance of the models through competitive game play, show which best inform strategy, and demonstrate statistical significance of the improvement over a simple strategy. Through this empirical study, we indicate features that we expect to be helpful in future improvements to Gin Rummy play.
Inspired by explanations of machine learning concepts in children’s books, we developed an approach to introduce supervised, unsupervised, and reinforcement learning using a block-based programming language in combination with the benefits of educational robotics. Instead of using blocks as high-end APIs to access AI cloud services or to reproduce the machine learning algorithms, we use them as a means to put the student “in the algorithm’s shoes.” We adapt the training of neural networks, Q-learning, and k-means algorithms to a design and format suitable for children and equip the students with hands-on tools for playful experimentation. The children learn about direct supervision by modifying the weights in the neural networks and immediately observing the effects on the simulated robot. Following the ideas of constructionism, they experience how the algorithms and underlying machine learning concepts work in practice. We conducted and evaluated this approach with students in primary, middle, and high school. All the age groups perceived the topics to be very easy to moderately hard to grasp. Younger students experienced direct supervision as challenging, whereas they found Q-learning and k-means algorithms much more accessible. Most high-school students could cope with all the topics without particular difficulties.
The industry demand for AI experts raised to unprecedented levels in the last years. However, the increasing demand was not met by the number of skilled professionals in this area. As an effort to mitigate this problem, many companies create AI residency programs to provide in-house practical training. However, we argue that the usual dynamics based on one-on-one mentorship in those programs is very hard to scale and insufficient to meet the demand for AI professionals. In this paper, we describe a hybrid AI residency program that connects educational institutions, partner companies, and prospective residents. This program is designed to be funded by partner companies.Residents are exposed to practical projects of industry interest and are instructed on AI techniques and tools. We describe how we implemented our program, the challenges involved, and the lessons learned after the conclusion of the first residency class. Our program was developed to be inclusive and scalable, and resulted in a high employment rate for our alumni. Furthermore, several partner companies invested in in-house AI teams after the residency, resulting in direct benefits for our local AI community.
We develop a data-driven approach for hand strength evaluation in the game of Gin Rummy. Employing Convolutional Neural Networks, Monte Carlo simulation, and Bayesian reasoning, we compute both offensive and defensive scores of a game state. After only one training cycle, the model was able to make sophisticated and human-like decisions with a 55.4% +/- 0.8% win rate (90% confidence level) against a Simple player.
With children talking to smart-speakers, smart-phones and even smart-microwaves daily, it is increasingly important to educate students on how these agents work—from underlying mechanisms to societal implications. Researchers are developing tools and curriculum to teach K-12 students broadly about artificial intelligence (AI); however, few studies have evaluated these tools with respect to AI-specific learning outcomes, and even fewer have addressed student learning about AI-based conversational agents. We evaluated our Conversational Agent Interface for MIT App Inventor and workshop curriculum with respect to 8 AI competencies from the literature. Furthermore, we analyze teacher (n=9) and student (n=47) feedback from workshops with the interface and recommend that future work (1) leverages design considerations to optimize engagement, (2) collaborates with teachers, and (3) addresses a range of student abilities through pacing and opportunities for extension. We found evidence for student understanding of all 8 competencies, with the most difficult concepts being AI ethics and machine learning. We recommend emphasizing these topics in future curricula.