Blackjack¶

Blackjack is a globally popular banking game known as Twenty-One. Theobjective is to beat the dealer by reaching a higher score than thedealer without exceeding 21. In the toolkit, we implement a simpleversion of Blackjack. In each round, the player only has two options:“hit” which will take a card, and ‘stand’ which end the turn. The playerwill “bust” if his hands exceed 21 points. After the player completeshis hands (chooses “stand” and has not busted), the dealer then realshis hidden card and “hit” until obtaining at least 17 points.

Leduc Hold’em is a simplified version of Texas Hold’em. Rules can be found here.

State Representation of Blackjack¶

1. Leduc Hold’em is a two player poker game. The deck used in Leduc Hold’em contains six cards, two jacks, two queens and two kings, and is shuffled prior to playing a hand. At the beginning of a hand, each player pays a one chip ante to the pot and receives one private card. A round of betting then takes place starting with player one.
2. Reinforcement Learning / AI Bots in Card (Poker) Games - Blackjack, Leduc, Texas, DouDizhu, Mahjong, UNO. JonathanLehner/rlcard.

In this toy environment, we encode the stateas an array [player_score,dealer_score] where player_score isthe score currently obtained by the player, and the dealer_score isderived from the card that faces up from the dealer.

Action Encoding of Blackjack¶

There are two actions in the simple Blackjack. They areencoded as follows:

Payoff of Blackjack¶

The player may receive a reward -1 (lose), 0 (tie), or 1 (win) in theend of the game.

Leduc Hold’em¶

Leduc Hold’em is a smaller version of Limit Texas Hold’em (firstintroduced in Bayes’ Bluff: Opponent Modeling inPoker). The deckconsists only two pairs of King, Queen and Jack, six cards in total.Each game is fixed with two players, two rounds, two-bet maximum andraise amounts of 2 and 4 in the first and second round. In the firstround, each player puts 1 unit in the pot and is dealt one card, thenstarts betting. In the second round, one public card is revealed first,then the players bet again. Finally, the player whose hand has the samerank as the public card is the winner. If neither, then the one withhigher rank wins. Other rules such as ‘fold’ can refer to Limit Texashold’em.

State Representation of Leduc Hold’em¶

The state is encoded as a vector of length 36. The first 3 elementscorrespond to hand card. The next 3 elements correspond to public card.The last 30 elements correspond the chips of the current player and theopponent (the chips could be in range 0~14) The correspondence betweenthe index and the card is as below.

Action Encoding of Leduc Hold’em¶

The action encoding is the same as Limit Hold’em game.

Payoff of Leduc Hold’em¶

The payoff is calculated similarly with Limit Hold’em game.

Limit Texas Hold’em¶

Texas Hold’em is a popular betting game. Each player is dealt twoface-down cards, called hole cards. Then 5 community cards are dealt inthree stages (the flop, the turn and the river). Each player seeks thefive best cards among the hole cards and community cards. There are 4betting rounds. During each round each player can choose “call”,“check”, “raise”, or “fold”.

In fixed limit Texas Hold’em. Each player can only choose a fixed amountof raise. And in each round the number of raises is limited to 4.

State Representation of Limit Texas Hold’em¶

The state is encoded as a vector of length 72. The first 52 elementsrepresent cards, where each element corresponds to one card. The hand isrepresented as the two hole cards plus the observed community cards sofar. The last 20 elements are the betting history. The correspondencebetween the index and the card is as below.

Action Encoding of Limit Texas Hold’em¶

There 4 actions in Limit Texas Hold’em. They are encoded as below.

Payoff of Limit Texas Hold’em¶

The stardard unit used in the leterature is milli big blinds per hand(mbb/h). In the toolkit, the reward is calculated based on big blindsper hand. For example, a reward of 0.5 (-0.5) means that the player wins(loses) 0.5 times of the amount of big blind.

Dou Dizhu¶

Doudizhu is one of the most popular Chinese card games with hundreds ofmillions of players. It is played by three people with one pack of 54cards including a red joker and a black joker. After bidding, one playerwould be the “landlord” who can get an extra three cards, and the othertwo would be “peasants” who work together to fight against the landlord.In each round of the game, the starting player must play a card or acombination, and the other two players can decide whether to follow or“pass.” A round is finished if two consecutive players choose “pass.”The player who played the cards with the highest rank will be the firstto play in the next round. The objective of the game is to be the firstplayer to get rid of all the cards in hand. For detailed rules, pleaserefer to Wikipedia orBaike.

In the toolkit, we implement a standard version of Doudizhu. In thebidding phase, we heuristically designate one of the players as the“landlord.” Specifically, we count the number of key cards orcombinations (high-rank cards and bombs), and the player with the mostpowerful hand is chosen as “lanlord.”

State Representation of Dou Dizhu¶

At each decision point of the game, the corresponding player will beable to observe the current state (or information set in imperfectinformation game). The state consists of all the information that theplayer can observe from his view. We encode the information into areadable Python dictionary. The following table shows the structure ofthe state:

State Encoding of Dou Dizhu¶

In Dou Dizhu environment, we encode the state into 6 feature planes. Thesize of each plane is 5*15. Each entry of a plane can be either 1 or 0.The 5 rows represent 0, 1, 2, 3, 4 corresonding cards, respectively. The15 columns start from “3” to “RJ” (Black Jack). For example, if we havea “3”, then the entry (1, 0) would be 1, and the rest of column 0 wouldbe 0. If we have a pair of “4”, then the entry (2, 1) would be 1, andthe rest of column 1 would be 0. Note that the current encoding methodis just an example to show how the feature can be encoded. Users areencouraged to encode the state for their own purposes by modifyingextract_state function in`rlcard/envs/doudizhu.py`__. The exampleencoded planes are as below:

Action Abstraction of Dou Dizhu¶

The size of the action space of Dou Dizhu is 27472. This number is toolarge for learning algorithms. Thus, we make abstractions to theoriginal action space and obtain 309 actions. We note that some recentstudies also use similar abstraction techniques. The main idea of theabstraction is to make the kicker fuzzy and only focus on the major partof the combination. For example, “33344” is abstracted as “333 **”.When the predicted action of the agent is not legal, the agent willchoose “pass.”. Thus, the current environment is simple, since oncethe agent learns how to play legal actions, it can beat random agents.Users can also encode the actions for their own purposes (such asincreasing the difficulty of the environment) by modifyingdecode_action function inrlcard/envs/doudizhu.py. Users are alsoencouraged to include rule-based agents as opponents. The abstractionsin the environment are as below. The detailed mapping of action and itsID is inrlcard/games/doudizhu/jsondata/action_space.json:

Payoff¶

If the landlord first get rid of all the cards in his hand, he will winand receive a reward 1. The two peasants will lose and receive a reward0. Similarly, if one of the peasant first get rid of all the cards inhand, both peasants will win and receive a reward 1. The landlord willlose and receive a reward 0.

Leduc Holdem Game

Simple Dou Dizhu¶

Simple Dou Dizhu is a smaller version of Dou Dizhu. The deck onlyconsists of 6 ranks from ‘8’ to ‘A’ (8, 9, T, J, Q, K, A), there arefour cards with different suits in each rank. What’s more, unlikelandlord in Dou Dizhu, the landlord in Simple Dou Dizhu only has onemore card than the peasants. The rules of this game is the same as therules of Dou Dizhu. Just because each player gets fewer cards, they endthe game faster.

State Representation of Simple Dou Dizhu¶

This is almost the smae as the state representation of Dou Dizhu, butthe number of the ‘deck’ has reduced from 54 to 28, and the number ofthe ‘seen cards’ reduced from 3 to 1. The following table shows thestructure of the state:

State Encoding of Simple Dou Dizhu¶

Leduc Hold'em Rules

The state encoding is the same as Dou Dizhu game.

Action Encoding of Simple Dou Dizhu¶

The action encoding is the same as Dou Dizhu game. Because of thereduction of deck, the actions encoded have also reduced from 309 to131.

Payoff of Simple Dou Dizhu¶

The payoff is the same as Dou Dizhu game.

Mahjong¶

throughout the world since 20th century. It is commonly played by 4players. The game is played with a set of 136 tiles. In turn playersdraw and discard tiles until The goal of the game is to complete the leagal hand using the 14thdrawn tile to form 4 sets and a pair. We revised the game into a simpleversion that all of the winning set are equal, and player will win aslong as she complete forming 4 sets and a pair. Please refer the detailon Wikipedia orBaike.

State Representation of Mahjong¶

The state representation of Mahjong is encoded as 6 feature planes,where each plane has 34 X 4 dimensions. For each plane, the column ofthe plane indicates the number of the cards in the given cards set, andthe row of the plane represents each kind of cards (Please refer to theaction space table). The information that has been encoded can berefered as follows:

Action Space of Mahjong¶

There are 38 actions in Mahjong.

Payoff of Mahjong¶

The reward is calculated by the terminal state of the game, wherewinning player is awarded as 1, losing players are punished as -1. Andif no one win the game, then all players’ reward will be 0.

No-limit Texas Hold’em¶

No-limit Texas Hold’em has similar rule with Limit Texas Hold’em. Butunlike in Limit Texas Hold’em game in which each player can only choosea fixed amount of raise and the number of raises is limited. In No-limitTexas Hold’em, The player may raise with at least the same amount asprevious raised amount in the same round (or the minimum raise amountset before the game if none has raised), and up to the player’sremaining stack. The number of raises is also unlimited.

State Representation of No-Limit Texas Hold’em¶

The state representation is similar to Limit Hold’em game. The state isrepresented as 52 cards and 2 elements of the chips of the players asbelow:

Action Encoding of No-Limit Texas Hold’em¶

There are 6 actions in No-limit Texas Hold’em. They are encoded asbelow.

*Note: Starting from Action ID 3, the action means the amount playershould put in the pot when chooses ‘Raise’. The action ID from 3 to 5corresponds to the bet amount from half amount of the pot, full amountof the pot to all in.

Payoff of No-Limit Texas Hold’em¶

The reward is calculated based on big blinds per hand. For example, areward of 0.5 (-0.5) means that the player wins (loses) 0.5 times of theamount of big blind.

UNO¶

Uno is an American shedding-type card game that is played with aspecially deck.The game is for 2-10 players. Every player starts withseven cards, and they are dealt face down. The rest of the cards areplaced in a Draw Pile face down. Next to the pile a space should bedesignated for a Discard Pile. The top card should be placed in theDiscard Pile, and the game begins. The first player is normally theplayer to the left of the dealer and gameplay usually follows aclockwise direction. Every player views his/her cards and tries to matchthe card in the Discard pile. Players have to match either by thenumber, color, or the symbol/action. If the player has no matches, theymust draw a card. If that card can be played, play it. Otherwise, keepthe card. The objective of the game is to be the first player to get ridof all the cards in hand. For detailed rules, please refer toWikipedia or UnoRules. And in our toolkit, the number ofplayers is 2.

State Representation of Uno¶

In state representation, each card is represented as a string of colorand trait(number, symbol/action). ‘r’, ‘b’, ‘y’, ‘g’ represent red,blue, yellow and green respectively. And at each decision point of thegame, the corresponding player will be able to observe the current state(or information set in imperfect information game). The state consistsof all the information that the player can observe from his view. Weencode the information into a readable Python dictionary. The followingtable shows the structure of the state:

State Encoding of Uno¶

In Uno environment, we encode the state into 7 feature planes. The sizeof each plane is 4*15. Row number 4 means four colors. Column number 15means 10 number cards from 0 to 9 and 5 special cards—“Wild”, “Wild DrawFour”, “Skip”, “Draw Two”, and “Reverse”. Each entry of a plane can beeither 1 or 0. Note that the current encoding method is just an exampleto show how the feature can be encoded. Users are encouraged to encodethe state for their own purposes by modifying extract_state functionin `rlcard/envs/uno.py`__. The example encodedplanes are as below:

We use 3 planes to represnt players’ hand. Specifically, planes 0-2represent 0 card, 1 card, 2 cards, respectively. Planes 4-6 are thesame.

Action Encoding of Uno¶

There are 61 actions in Uno. They are encoded as below. The detailedmapping of action and its ID is inrlcard/games/uno/jsondata/action_space.json:

Payoff of Uno¶

Each player will receive a reward -1 (lose) or 1 (win) in the end of thegame.

Gin Rummy¶

Gin Rummy is a popular two person card game using a regular 52 card deck(ace being low). The dealer deals 11 cards to his opponent and 10 cardsto himself. Each player tries to form melds of 3+ cards of the same rankor 3+ cards of the same suit in sequence. If the deadwood count of thenon-melded cards is 10 or less, the player can knock. If all cards canbe melded, the player can gin. Please refer the detail onWikipedia.

If a player knocks or gins, the hand ends, each player put down theirmelds, and their scores are determined. If a player knocks, the opponentcan layoff some of his deadwood cards if they extend melds of theknocker. The score is the difference between the two deadwood counts ofthe players. If the score is positive, the player going out receives it.Otherwise, if the score is zero or negative, the opponent has undercutthe player going out and receives the value of the score plus a 25 pointundercut bonus.

The non-dealer discards first (or knocks or gins if he can). If theplayer has not knocked or ginned, the next player can pick up thediscard or draw a card from the face down stockpile. He can knock or ginand the hand ends. Otherwise, he must discard and the next playercontinues in the same fashion. If the stockpile is reduced to two cardsonly, then the hand is declared dead and no points are scored.

State Representation of Gin Rummy¶

The state representation of Gin Rummy is encoded as 5 feature planes,where each plane is of dimension 52. For each plane, the column of theplane indicates the presence of the card (ordered from AS to KC). Theinformation that has been encoded can be referred as follows:

Action Space of Gin Rummy¶

There are 110 actions in Gin Rummy.

Payoff of Gin Rummy¶

The reward is calculated by the terminal state of the game. Note thatthe reward is different from that of the standard game. A player whogins is awarded 1 point. A player who knocks is awarded 0.2 points. Thelosing player is punished by the negative of their deadwood countdivided by 100.

Leduc Hold'em Card Game

If the hand is declared dead, both players are punished by the negativeof their deadwood count divided by 100.

Settings¶

The following options can be set.

Variations¶

One can create variations that are easier to train by changing theoptions and specifying different scoring methods.

Fall 2020

Public Reports

Public Video Reports

Other Titles

• Tic Tac Toe Solver
• Beating the house in Blackjack
• Effect of Noise on Learning a Planar Pushing Task using SAC
• ResQNet: Finding Optimal Fire Rescue Routes
• COVID Chatbot
• Regularized Follow-the-Leader in Online MDP for Efficient Topographical Mapping
• Learning POMDP model parameters from missing observations
• Reinforcement Learning for No-Limit Texas Hold ‘Em with Bomb Pots
• Identifying Optimal Locations for Satellite Image Capture
• Diet Conscious Meal Planner
• Mitigating Risk of Public Transit during COVID-19
• Predicting the Match: Using Bayesian Networks to Predict Professional Tennis Outcomes
• Efficient Single-Agent Capture of a Moving Target
• Q-Learning Applied to the Taxi Problem
• Settlers of CATAN
• Autonomous Snake
• Q-Learning for Pre-Flop Texas Hold ‘Em
• Deep RL for Atari Games
• Simulating a D&D Encounter with Q-Learning
• Deep RL for Automated Stock Trading
• Dating Under Uncertainty
• Retinal Implant Electrical Stimulation via RL
• Batch Offline RL in the Healthcare Setting
• Computer Caddy – Using RL to advice Golfers’ Club Selection
• RL for Fischer Random Chess
• Timely Decision Making with Probability Path Model
• Satellite-Imagery Based Poverty Level Evaluation System in Mexico with Deep RL Approach
• Pokemon Showdown
• Deep RL for Space Invaders
• Learning to Run
• RL-Based Control of Policy Selection in Near-Accident Scenarios
• Model Predictive Control for an Aircraft Autopilot
• Finding Inharmonic Timbres Locally Consonant with Arbitrary Scales
• Escape Roomba
• Driving in Traffic
• Playing Snake
• The 2020 FLatland Challenge
• Elevator Scheduling with Neural Q-Learning
• Optimizing Immunotherapy Treatment using RL
• Modeling Leduc Hold ‘Em Poker
• Auto Trading System Using Q-Learning
• Energy System Modeling
• Optimizing Fox in the Forest through RL
• Learning Gin Rummy
• Car Racing with Deep RL
• Sequential Decision Making for Mineral Exploration
• Advanced Driver Assistance Systems
• Learning to Play Stargunner with Deep Q-Networks
• A Fourth-and-Goal Football Recommender System
• Algorithms for Motion Planning
• Playing Farkle
• Connect 4: A Survey of Different RL Techniques to Destroy Your Pride
• Decision Making in the word game, Codenames
• Reinforcement Learning Approaches for An Adversarial Snake Agent
• An Attention-Based, Reinforcement-Learned Heuristic Solver for the Double Travelling Salesman Problem With Multiple Stacks
• Uncertainty Aware Model-Based Policy Optimization
• Navigating the Four-Way Stop Autonomously
• Ground Water Remediation Using Sequential Decision Making
• Final Project: Satellite Collision Avoidance
• Q Learning for 4th Down Decision Making in the NFL
• Q-Learning for the Game of Nim: Does The Agent Learn a Combinatorially Optimal Strategy On Its Own?
• Contextual Bandit Algorithms in Recommender Systems
• A Comparison of Reinforcement Learning Methods for Autonomous Navigation
• Reinforcement Learning for Behavior Planning in Intersections for Autonomous Vehicles
• Reinforcement Learning for Pacman Capture the Flag
• Comparing Different Optimization Techniques for Learning Continuous Control with Neural Networks
• Autonomous Exploration in Subterranean Environments
• Improving Image Denoising through Decision Making
• Using MDPs to Optimally Allocate Funds
• Explanations Meet Decision Theory
• Learning Policies for Adaptive LiDAR Scanning with POMDPs
• Cautious Markov Games: A New Framework for Human-Robot Interaction
• Selecting a multibasis community structure for the connectome
• Reinforcement Learning Techniques for Long-Term Trading and Portfolio Management
• Optimal Asset Allocation with Markov Decision Processes
• Symbolic Regression with Bayesian Networks
• Scheduling battery charging using deep reinforcement learning
• Online Knapsack Problem Using Reinforcement Learning
• Policy gradient optimization for
• Resource Allocation for Wildfire Prevention
• Using Reinforcement Learning to Play Omaha
• Fraud Detection for Mobile Payments using Bayesian Network and CNN
• Neuro-Adaptive Artificial Neural Networks for Reinforcement Learning
• AI Agent for Qwirkle
• Learning Optimal Wildfire Suppression Policies With Reinforcement Learning
• Bid Smart with Uncertainty: An Autonomous Bidder
• AA228/CS238 Final Report
• Modeling Identification of Approaching Aircraft as a POMDP
• Short-Term Trading Policies for Bitcoin Cryptocurrency Using Q-learning
• Reinforcement Learning of a Battery Power Schedule for a Short-Haul Hybrid-Electric Aircraft Mission
• Autonomous Helicopter Control for Rocket Recovery
• Reinforcement Learning Strategies Solving Game Gomoku
• A Wildfire Evacuation Recommendation System
• Battleship with Alogrithm
• Developing an Optimal Structure for Breast Cancer Single Cell Classification
• Utilizing Deep Q Networks to Optimally Execute Stock Market Entrance and Exit Strategies
• Online Planning for a Grid World POMDP
• Contingency Manager Agent for Safe UAV Autonomous Operations
• Solving Mastermind as a POMDP
• Simulated Drone Flight with Advantaged Actor Critic Reinforced Learning in 2 and 3 Dimensions
• Solving Queueing Problem Using Monte Carlo Tree Search
• Bayesian Structure Learning on NFL play data
• Multi-Agent Rendezvous Using Reinforcement Learning
• Dynamic Portfolio Optimization
• Fairness and Efficiency in Multi-Portfolio Liquidation: An Multiple-Agent Deep Reinforcement Learning Approach
• Evaluating Poker Hands
• Saving Artificial Intelligence Clinician
• Evaluation of online trajectory planning methods for autonomous vehicles
• Solving Leduc Hold’em Counterfactual Regret Minimization
• From aerospace guidance to COVID-19: Tutorial for the application of the Kalman filter to track COVID-19
• A Reinforcement Learning Algorithm for Recycling Plants
• Monte Carlo Tree Search with Repetitive Self-Play for Tic-Tac-Toe
• Developing a Decision Making Agent to Play RISK

Fall 2019

Public Reports

Leduc Hold'em Strategy

Other Titles

• Linear Array Target Motion Analysis Using POMDPs
• Speed or Safety?: Calculating Urban Walking Routes Based on Probability of Crime and Foot Traffic
• AlphaGomoku
• Modelling Uncertainty in Dynamic Real-time Multimodal Routing Problems
• Reinforcement Learning for Portfolio Allocation
• Preparation of Papers for AIAA Technical Conferences
• Autonomous Racing
• Deep Learning Enabled Uncorrelated Space Observation Association
• Landing a Lunar Spacecraft with Deep Q-Learning
• POkerMDP: Decision Making for Poker
• 1V1 Leduc Hold’em Bot
• Political Influencers: Using Election Finance Data to Analyze Campaign Success via Bayesian Networks
• Developing AI Policies for Street Fighter via Q-learning
• Impact of Market Technical Indicators On Future Stock Prices Using Reinforcement Learning
• Allocation of Hearts for Transplant as an MDP
• Multi-Agent Reinforcement Learning in a 2D Environment for Transportation Optimization
• Planning under Uncertainty for Discrete Robotic Navigation with Partial Observability
• Deep Reinforcement Learning Applied to Mid-Frequency Trading
• Application of Subspace Identification for Classification of Neural-Activity during Decision-Making
• Using Markov-Decision Processes to Design Betting Strategies for the NFL
• Maneuvering Characteristics Control Systems using Discrete-Time MDPs
• MDP Based Motion Planning In Unsignaled Intersections
• Competitive Blackjack Using Reinforcement Learning
• Modelling Pedestrian Vehicle Interaction at Stop Sign using Markov Decision Process
• Jeopardy! Wagering Under Uncertainty
• Love Letters Under Uncertainty
• Playing The Resistance with a POMDP
• Robotic Simultaneous Localization and Mapping with 2D Laser Scan
• Mars Rover: Navigating an Uncertain World
• Modeling Blood Donations Over Time as a POMDP
• Reinforcement Learning for Control on OpenAI Gym Environments
• Playing Connect 4 using Reinforcement Learning
• Evaluation of Reduced Algorithmic Complexity for Grasping Tasks by Using a Novel Underactuated Curling Grasper with Reinforcement Learning
• Optimizing Strategies for Settlers of Catan
• Exploring Search Algorithms for Klondike Solitaire
• A Sparse Sampling Control Strategy for Risk Minimization during Stretchable Sensor Network Deployment
• Computing Strategies for the 7 Wonders Board Game
• POMDP modeling of stochastic Tetris
• Solving a Maze with Doors and Hidden Tigers
• Playing “Dominion” with Deep Reinforcement Learning
• Delivery Vehicle Navigation in Crowd with Reinforcement Learning
• Capturing Uncertainty in a Multi-Modal Setting With JRMOT: A Real-Time 2D-3D Multi-Object Tracker
• Decentralized Satellite Network Communication
• Seismic Network Planning
• Reinforcement Learning for PaoDeKuai, A Card Game
• Training A Bai Fen Agent with Reinforcement Learning
• Decision Making for Launch Cancellation Based Upon Storm Conditions
• Optimizing for the Competitive Edge: Modeling Sequential Binary Decision Making for Two Competing Firms
• Datacenter Equipment Maintenance Optimization
• To Heat Or Not To Heat: Reinforcement Learning for Optimal Residential Water Heater Control
• Learning to Play Snake Game with Deep Reinforcement Learning
• Optimal Traffic Light Control for Efficient City Transportation
• Modeling NBA Point Spread Betting as an MDP
• Solving a car racing game using Reinforcement Learning
• Is Uncertainty Really Harmful: Solving Partially Observable Lunar Lander Problem with Deep Reinforcement Learning
• Autonomous Navigation of an RC Boat Under a POMDP Framework
• Evaluating the Bayes-Adaptive MDP Framework on Stochastic Gridworld Environments
• Value Iteration with Enhanced Action Space for Path Planning
• The Medical Triage Problem: Improving Hospitals’ Admission Decisions
• Optimal Route Selection for Riders in Toronto
• Model Free Learning for Optimal Power Grid Management
• Wasting Less Time on the Road Using MDPs
• Learning User Preferences to Produce Sequential Movie Recommendations
• A Comparison of Learning Based Control Methods for Optimal Trajectory Tracking with a Quadrotor
• Artificial Pancreas: Q-Learning Based Control for Closed-loop Insulin Delivery Systems
• Navigating in an Uncertain World
• Teaching an Autonomous Car to Drive through an Intersection with POMDPs
• Atomic structure minimization using simulated annealing with a MCTS temperature scheme
• AI Game Player for 2048
• Deep Q-Learning with GARCH for Stock Volatility Trading
• Learning to Become President
• Solving GNSS Integrity-Based Path Planning in Urban Environments via a POMDP Framework
• Reservoir operation under climate uncertainty
• Reinforcement Learning for Maze Solving
• Using Reinforcement Learning to Find Basins of Attraction
• Planning for Asteroid Prospecting Missions with POMDPs
• Human-Aware Robot Motion Planner
• Determining Federal Funds Rate Changes – Hike / Cut / Hold – Under Economic Uncertainty
• Simulating Work-Life Balance with POMDP
• Solving 2048 as a Markov Decision Process
• Accounting for Delay in Dynamic Resource Allocation for Wildfire Suppression – a POMDP Approach
• Daily Allocation of Assets with Distinct Risk Profiles using Reinforcement Learning
• LocoNets for Deep Reinforcement Learning
• Exploring a full joint observability game with Markov decision processes
• Deep Bayesian Active Learning for Multiple Correct Outputs
• Convolutionally Reducing Markov Decision Processes
• Robust Decision Making Agent for Frozen Lake
• Tic-Tac-Toe How Many In A Row?
• Turbomachinery Optimization Under Uncertainty
• Devising a Policy for Liar’s Dice Using Model Free Reinforcement Learning
• Political Compromises: an Iterative Game of Prisoner’s Dilemma
• Optimal Home Energy System Management using Reinforcement Learning
• Drone Tracking in a 2-dimentional Grid using Particle Filter Algorithm
• Deep Reinforcement Learning for Traffic Signal Control
• A Deep Reinforcement Learning Approach to Recommender Systems
• FlyCroTugs – Collaborative Object Manipulation Using Flying Tugs
• Local Approximation Q-Learning for a Simplified Satellite Reconnaissance Mission
• Developing Policies for Blackjack Using Reinforcement Learning
• Applying Q-learning to the Homicidal Chauffeur Problem
• Optimal Satellite Detumbling through Reinforcement Learning
• Active Preference-Based Gaussian Process Regression for Reward Learning and Optimization
• A Comparative Study on Heart Disease Prediction
• Robot Navigation with Human Intent Uncertainty
• Conquering the Queen of Spades: A Hearts Agent
• Using Markov Decision Processes to Predict Soccer Player Market Value
• Effectiveness of Recurrent Network for Partially-Observable MDPs
• Capture The Flag
• Predicting uncertainty
• Optimal Asset Allocation with Markov Decision Processes
• Nets on Nets: Using Bayesian Networks to Predict Supplier Links in Economic Networks
• Playing 2048 With Reinforcement Learning
• Trading strategies using deep reinforcement learning with news and time series stock data
• Modeling Contract Bridge as a POMDP
• Solving Rubik’s Cubes Using Milestones
• Playing 2048 with Deep Reinforcement Learning
• An Approximate Dynamic Programming Minimum-Time Guidance Policy for High Altitude Balloons
• Identifying Bots on Twitter
• Approaches to Model-Free Blackjack
• Jumping Robot Simulator: An Exploration of Methods to Teach a Bio-Inspired Frog Robot to Navigate
• Air Traffic Control Tower Policy for Terminal Environment Operations
• Managing a Prediction Market Portfolio
• Applying Partially Observable Markov Decision Making Processes to a Product Recommendation System
• Self-Driving Under Uncertainty
• Reinforcement Learning for QWOP
• Modeling Macroeconomic Phenomena with Multi-Agent Reinforcement Learning
• Optimal Learning Policy via POMDP planning
• AI Guidance for Thermalling in a Glider
• Decision Making For Profit: Portfolio Management using Deep Reinforcement Learning
• Self-play Reinforcement Learning for Open-face Chinese Poker
• Feature Constrained Graph Generation with a Modified Multi-Kernel Kronecker Model
• Sensor Fusion of IMU and LiDAR Data Using a Multirate Extended Kalman Filter
• Optimizing Empiric Antibiotic Delivery in the Emergency Department
• The Task Completion Game
• Optimizing Modified Mini-Metro (M³)
• Improving Pragmatic Inferences with BERT and Rational Speech Act Framework and Data Augmentation
• Deep Q-Learning for Playing Hanabi as a POMDP
• A Comparative Study on Heart Disease Prediction

Fall 2018

Public Reports

Other Titles (excluding optional final projects)

• Occlusion Handling for Local Path Planning with Stereo Vision
• Pre-Flop Betting Policy in Poker
• Optimal Impulsive Maneuver Times for Simultaneous Imaging and Gravity Recovery of an Asteroid
• Monte-Carlo Planning in Subsurface Resource Development
• Learning to Win at Go-Stop
• Police Officer Distribution
• Optimizing Road Construction to Improve Traffic Conditions Using Reinforcement Learning
• Q-Learning for Casino Hold’em
• Modeling a Connected Highway Merge as a POMDP Using Dynamic GPS Error
• Figure 8 Race Track Optimal and Safe Driving
• Predictive Maintenance of Trucks using POMDPs
• Predictive Models for Maximizing Return on Agriculture given Location and Temperature
• A Policy to Deal With Delay Uncertainty
• Reinforcement Learning Methods for Energy Microgrids
• Boom! Tetris for Bot – Designing a Reinforcement Learning Framework for NES Tetris
• Hidden Markov Models for Economic Cycle Regime Estimation
• Push Me: Optimizing Notification Timing to Promote Physical Activity
• Resource Allocation for Floridian Hurricanes
• Motion Planning in Human-Robot Interaction Using Reinforcement Learning
• Automated Neural Network Architecture Tuning with Reinforcement Learning
• Imitation Learning in OpenAI Gym with Reward Shaping
• Collision Avoidance for Unmanned Rockets using Markov Decision Processes
• MDP Solvers for a Successful Sushi Go! Agent
• Uncovering Personalized Mammography Screening Recommendations through the use of POMDP Methods
• Implementing Particle Filters for Human Tracking
• Decision Making in the Stock Market: Can Irrationality be Mathematically Modelled?
• Single and Multi-Agent Autonomous Driving using Value Iteration and Deep Q-Learning
• Buying and Selling Stock with Q-Learning
• Application and Analysis of Online, Offline, and Deep Reinforcement Learning Algorithms on Real-World Partially-Observable Markov Decision Processes
• Reward Augmentation to Model Emergent Properties of Human Driving Behavior Using Imitation Learning
• Classification and Segmentation of Cancer Under Uncertainty
• Comparison of Learning Methods for Price Setting of Airfare
• QMDP Method Comparisons for POMDP Pathfinding
• Global Value Function Approximation using Matrix Completion
• Artificial Intelligence Techniques for a Game of 2048
• Exploring the Boundaries of Art
• An Iterative Linear Algebra Approach to Dynamic Programming
• Solving Open AI Gym’s Lunar Lander with Deep Reinforcement Learning
• Application of Imitation Learning to Modeling Driver Behavior in Generalized Environments
• Craps Shoot: Beating the House…?
• Movie Recommendations with Reinforcement Learning
• Playing Atari 2600 Games Using Deep Learning
• Traverse Synthesis for Planetary Exploration
• Optimal operation of an islanded microgird under a Markov Decision Process framework
• Implementing Deep Q-learning Extensions in Julia with Flux.jl
• Learning How to Buy Food
• Using Dynamic Programming for Optimal Meal Planning
• Modelling Wildfire Evacuation using MDPs
• Comparing Multimodal Representations for Robotic Reinforcement Learning Tasks
• Applying Reinforcement Learning to Packet Routing in Mesh Networks
• Xs & Os: Creating a Tic-Tac-Toe Foe
• Doggo Does a Backflip: Deep Reinforcement Learning on a Quadruped Platform
• GrocerAI: Using Reinforcement Learning to Optimize Supermarket Purchases
• Reinforcement Learning For The Buying and Selling of Financial Assets
• Towards Designing a Policy on Automotive GPS Integrity
• Generalized Kinetic Component Analysis
• Trading Wheat Futures Contracts
• Using PCR, Neural Networks, and Reinforcement Learning
• Reinforcement Learning for Inverted Pendulums
• Electric Vehicle Charging under Uncertainty
• Automatic Accompaniment Generator: An MDP Approach
• Comparison of Methods in Artificial Life
• Modeling a Better Visual Acuity Test
• Online Methods Applied to the Game of Euchre
• Missile Defense Strategy: Towards Optimal Interceptor Allocation
• Smart Charging of Electric Vehicles under State Uncertainty
• Learning to Play Atari Breakout Using Deep Q-Learning and Variants
• Decision making on fault-code
• Learning FlappyBird with Deep Q-Networks
• A Fresh Start: Using Reinforcement Learning to Minimize Food Waste and Stock-Outs in Supermarkets
• Autonomous orbital maneuvering using reinforcement learning
• Autonomous Decision-Making for Space Debris Avoidance
• Maximizing Monthly Expenditures Under Uncertainty
• Modeling Voter Preferences in US General Elections
• Application of Reinforcement Learning to the Path Planning with Dynamic Obstacles
• A Decision Making framework for Medical Diagnostics
• Learning to Walk Using Deep RL
• Q(λ)-Learning with Boltzmann and ε-greedy Exploration Applied to a Race Car Simulation
• Reinforcement learning for Glassy/Phase Transitions
• Proximal Policy Optimization in Julia
• University Technology Patent and License Decisions: Open- versus Closed-Loop Planning in a Markov Decision Process
• A Policy Gradient Approach for Continuous-Time Control of Spacecraft Manipulator Systems
• Applying Techniques in Reinforcement Learning to Motion Planning in Redundant Robotic Manipulators
• Deep Q-Learning for Atari Pong
• Adversarial Curiosity for Model-Based Reinforcement Learning
• A Markov Decision Process Approach to Home Energy Management with Integrated Storage
• Using Maximum Likelihood Model-Based Reinforcement Learning to Play Skull
• Cryptocurrency Trading Strategy with Deep Reinforcement Learning
• Evaluating Multisense Word Embeddings Final Report
• Near-Earth Object (NEO) Deflection via POMDP
• Reinforcement Learning for Car Driving
• Reinforcement Learning for Automatic Wheel Alignment
• Julia Implementation of Trust Region Policy Optimization
• Deep Reinforcement Learning with Target and Prediction Networks
• Playing Tower Defense with Reinforcement Learning
• Q-Learning agent as a portfolio trader
• Multi-Robot Rendezvous from Indoor Acoustics
• Portfolio Asset Allocation using Reinforcement Learning
• Creating a 2048 AI Solver using Expectimax
• Robustness of Reinforcement Learning Based Communication Networks in Multi-Modal Multi-Step Games to Input Based Adversarial Attacks
• Deep Q-Learning with Nearest Neighbors in Sequential Decision-Making for Sepsis Treatment
• Positioning Archival Radar Data with a Particle Filter
• Reinforcement Learning for Atari Skiing
• Understanding Donations with Reinforcement Learning
• Known and Unknown Discrete Space Exploration Using Deep Q-Learning
• Speeding Up Reinforcement Learning with Imitation
• Learning Bandwidth-Limited Communication in Decentralized Multi-agent Reinforcement Learning

Fall 2017

• 2048 as a MDP
• A Computational Approach to Employee Resource Allocation between Multiple Projects
• Accelerated Training of Deep Q Learning Models for Atari Games
• AlphaOthello: Developing an Othello player through Reinforcement Learning on Deep Neural Networks
• An Online Approach to Energy Storage Management Optimization
• An Optimal Basketball Foul Strategy by Value Iteration
• Annealed Reward Functions in Continuous Control Reinforcement Learning
• Applications of Inverse Reinforcement Learning for Multi-Feature Path Planning
• Attributing Authorship in the Case of Anonymous Supreme Court Opinions Utilizing SVMs and Probabilistic Inference on Score Uncertainty
• Balancing Safety and Performance in Imitation Learning
• Baseball Pitch Calling as a Markov Decision Process
• Batch Reinforcement Learning Technological Investment Strategies Utilizing The Contingent Effectiveness Model In A Markov Decision Process
• Bayesian Learning of Image Transformations from User Preferences for Individualized Automatic Filters
• BetaMiniMax: An Agent for Cheat
• Building a Game Agent to Play Resistance
• Building Trust in Autonomy: Sharing Future Behavior in Reinforcement Learning
• Car racing with low dimensional input features
• Comparison of Classical Control Methods and POMDPs for 3D Motion Control
• Control of a Partially-Observable Linear Actuator
• DDQN Learning for 2048
• Deep Q-learning in OpenAI gym
• Deep Q-Learning with Target Networks for Game Playing
• Design of A Planning Machinery for Choosing an NBA Team’s Play Style Strategy
• Detecting Human from Image with Double DQN
• Determining the Optimal Betting Policy: World Series
• Disrupting Distributed Consensus (or Not) Using Reinforcement Learning
• Dominating Dominoes
• Double A3C: Deep Reinforcement Learning on OpenAI Gym Games
• Emergent Language in Multi-Agent Co-operative Reinforcement Learning
• Explore the Frontier of Safe Imitation Learning for Autonomous Driving
• Fast Operation of Coordinated Distributed Energy Resources without Network Models using Deep Deterministic Policy Gradients
• Faster Algorithms for Contextual Combinatorial Cascading Bandits
• Finding a Scent Source with a Soft Growing Robot Using Monte Carlo Tree Search
• Gaming Bitcoin Leveraging Model-Based Reinforcement Learning
• Get Ready for Demand Response
• GlideAI: Optimizing Soaring Strategy with Reinforcement Learning
• Grid Stability Management and Price Arbitrage for Distributed Energy Storage and Generation via Reinforcement Learning
• Guiding the management of sepsis with deep reinforcement learning
• HMMs for Prediction of High-Cost Failures
• Integrating Mini-Model Evidence into Policy Evaluation
• Investigating Parametric Insurance Models As Multi- Variable Decision Networks
• Learning an Optimal Policy for Police Resource Allocation on Freeways
• Learning Terminal Airspace Models from TSAA Data
• Learning the Education System
• Learning the Policy of the Policy: Deep Reinforcement Learning with Model-Based Feedback Controllers
• Learning to Play a Simplified Version of Monopoly Using Multi-Agent SARSA
• Learning to Play Othello Without Human Knowledge
• Limbed Robot Motion Control through Online Reinforcement Learning
• Linear Approximation Q-Learning to Learn Movement in a 2D Space
• Locally Optimal Risk Aware Path Planning
• Massively Parallel Reinforcement Learning Using Trust Region Policy Optimization
• Model-Free Learning of Casino Blackjack
• Model-Free Reinforcement Learning of a Modified Helicopter Game
• Model-Free Reinforcement Learning on Flappy Bird
• Modeling Disaster Evacuation Paths
• Modeling Flight Delay and Cancelation
• Modeling NBA Matchups
• Modeling Optimally Efficient Earth to Earth Flight Trajectories in Kerbal Space Program with Reinforcement Learning
• Modeling Real Estate Investment with Deep Reinforcement Learning
• Multi-Agent Cooperative Language Learning
• Multi-armed Bandits with Unobserved Confounders
• Multidisciplinary Design Optimization for Approximating Unsteady Aerodynamics of Flexible Aircraft Structures
• Navigating Chaos: Autonomous Driving in a Highly Stochastic Environment
• Optimal Flight Itineraries Under Uncertainty Using a Stochastic Markov Decision Process
• Optimal Strategy for Two-Player Incremental Classification Games Under Non-Traditional Reward Mechanisms
• Optimizing sequential time-lapse seismic davcx bta collection using a POMDP
• Personal Portfolio Asset Allocation as An MDP Problem
• Planetary Lander with Limited Sensor Information and Topographical Uncertainty
• Playing Flappy Bird Using Deep Reinforcement Learning
• POMDP and MDP for Underwater Navigation
• POMDP Modeling of a Simulated Automatic Faucet for Cognitive State and Task Recognition
• Portfolio Management
• Power Grid real time optimization using Q-Learning
• Predicting Congressional Voting Behavior and Party Affiliation using Machine Learning
• Predicting Income From OkCupid Profiles
• Predicting NBA Game Outcomes using POMDPs
• Predicting Subjective Sleep Quality
• Preparation of Papers for AIAA Technical Journals
• Pursuit-Evasion Game with an Agent Unaware of its Role
• Rapid Reinforcement Learning by Injecting Stochasticity into Bellman
• Real Time Collision Detection and Identification for Robotic Manipulators
• Reinforcement Learning Applied to Quadcopter Hovering
• Reinforcement Learning Approaches to Pathfinding in Stochastic Environments
• Reinforcement Learning For A Reach-Avoid Game
• Reinforcement Learning for Atari Breakout
• Reinforcement Learning for Crypto-Currency Arbitrage Bot
• Reinforcement Learning for Precision Landing of a Physically Simulated Rocket
• Reinforcement learning in an online multiplayer game
• Reinforcement Learning of Blackjack Variants
• Reinforcement training of nonlinear reduced order models
• Reward Shaping with Dynamic Guidance to Accelerate Learning for Multi-Agent Systems
• Risk – Bayesian World Conquest
• Roboat: Reinforcement of Boat’s Optimal Adaptive Trajectory
• Robotic Arm Motion Planning Based on Reinforcement Learning
• Robotic Decision Making Under Uncertainty
• Sensor Selection for Energy-Efficient Path Planning Under Uncertainty
• ShAIkespeare: Generating Poetry with Reinforcement Learning and Factor Graphs
• Shared Policies in Aircraft Avoidance
• Simulated Autonomous Driving with Deep Reinforcement Learning
• Simulating Coverage Path Planning with Roomba
• SLAMming into Obstacles: Simultaneous Localization and Mapping the Path of a Turtlebot
• Smart Health Coach: Using Markov Decision Processes to Optimize Health Advising Strategies
• Smarter Queues by Reinforcement Learning
• Solving Real-world Oil Drilling Problem with Multi-Armed Bandit and POMDP Models
• Stay in Your Lane: Probabilistic Vehicular Automation for DIY Robocars
• Supervised Learning and Reinforcement Learning for Algorithmic Trading
• Taking Out the GaRLbage
• Terrain Relative Navigation and Path Planning for Planetary Rovers
• Time-Constrained Sample Retrieval in a Martian Gridworld with Unknown Terrain
• Trade-offs in Connect Four Game-Playing Agents
• Training an Intelligent Driver on Highway Using Reinforcement Learning
• UAV Collision Avoidance Using Neural Network-Assisted Q-Learning
• Understanding Limitations of Network Meta-analysis Approaches to Rank Effectiveness of Treatments
• Using Bayesian Networks to Impute Missing Data
• Using Bayesian Networks to Predict Credit Card Default
• Using Bayesian Networks to Understand and Predict Wildfires
• Using Classification Models to Represent and Predict Students’ Restaurant Preferences
• Using Q-Learning to Optimize Lunar Lander Game Play
• Using the QMDP Method to Determine an Open Ocean Fishing Policy
• Utilizing fundamental factors in reinforcement learning for active portfolio management
• Utilizing Fundamental Factors in Reinforcement Learning for Active Portfolio Management

Fall 2016

• Model-Free Reinforcement Learning of Blackjack
• Partially Observable Actions in Solving Markov Decision Processes. The Case for Insulin Dosing Optimization in Diabetic Patients.
• Using Monte-Carlo Tree Search to Solve the Board Game Hive
• Blackjack: How to use MDP’s to (nearly) beat the house
• Cancer Metabolism Mapping: Bayesian Networks and Network Learning Techniques to Understand Cancer Metabolic and Regulatory Pathways
• Gibbs Sampling in BayesNets.jl
• UAV Collision Avoidance Policy Optimization with Deep Reinforcement Learning
• Improving Training Efficiency in Deep Q-Learning for Atari Breakout
• Monitoring Machine Workload Distribution with Kalman Filter
• Approximating Transition Functions to Cart Track MDPs via Sub-State Sampling
• Approaching Quantitative Trading with Machine Learning
• Structure and Parameter Learning in Bayesian Networks with Applications to Predicting Breast Cancer Tumor Malignancy in a Lower Dimension Feature Space
• Autonomous Racing by Learning from Professionals
• Bravo Zulu: Optimizing Teammate Selection for Military and Civilian Applications
• Investigating Transfer Learning in Deep Reinforcement Learning Context
• Simultaneous Estimation and Control with MCTS
• Controlling Soft Robots with POMCP
• Automatic Learning of Computer Users’ Habits
• Learning to Play Soccer in the OpenAI Gym
• Playing Ultimate Tic-Tac-Toe with TD Learning and Monte Carlo Tree Search
• A Bayesian Network Model of Pilot Response to TCAS Resolution Advisories
• Improving Head Impact Kinematics Measurement Accuracy using Sensor Fusion
• Drive Decision Making at Intersections
• Deterministic and Bayesian Techniques for Spaceborne Vision-Based Non-Stellar Object Detection
• A Two-Phased Deep Reinforcement Learning Algorith for High-Frequency Trading
• Implementation and Experimentation of a DQN solver in Julia for POMDPs.jl
• Landing on the Moon
• Deserted Island: Cooperative Behavior in Absence of Explicit Delayed Reward
• DeepGo.py
• Managing Groundwater under Uncertain Seasonal Recharge
• Using Reinforcement Learning to Find Flaws in Collision Avoidance Systems
• Effectiveness of Bayesian Networks in Building a Prediction Model for Movie Success
• Data Driven Agent based on Aircraft Intent
• Deep Q-Learning with Natural Gradients
• A Shot in the Dark: Beating Battleship with POMCP
• Accelerated Asynchronous Deep Reinforcement Learning Variant of Advantage Actor-Critic
• Applying Reinforcement Learning and Online Methods on the Inverted Pendulum Problem
• Predicting Sentiment with Deep Q-Learning
• A Lookahead Strategy for Super-Level Set Estimation using Gaussian Processes
• Modeling Breast Cancer Treatment as a Markov Decision Process
• Learning 31 using Cross-Entropy Methods
• Improving Haptic Guidance using Reinforcement Learning
• NLPLab: Actor-Critic Training in Natural Language Processing
• Deep Reinforcement Learning on Atari Breakout
• Reinforcement Learning for LunarLander
• Reinforcement Learning for AI Machine Playing Hearthstone
• Using Deep Q-Learning to Automate CNN Training
• Automatic Continuous Variable Encoder in Bayesian Network
• Side Channel Analysis using Neural Networks and Random Forests
• A Decision-Making System for Wildfire Management
• Decentralized Game Theoretic Methods for the Distributed Graph Coverage Problem
• Autonomous altitude control for high altitude balloons
• Neural Network Arbitration for Better Time and Accuracy trade-offs
• Deep Deterministic Policy Gradient with Robot Soccer
• Towards a Personal Decision Support System
• Optimal Gerrymandering under Uncertainty
• The Ambulance Dilemma: Crossing an Intersection with Monte Carlo Tree Search
• DeepDominionDevelopmental Policy Design: an MDP approach
• Training of a craps betting strategy with Reinforcement Learning Techniques
• Engineering a Better Monkey
• Decision Making During a Bicycle Race
• Using Discrete Pressure Measurements to Understand Subsonic Bluff-Body Dynamic Damping
• Effective Move Selection in Chess Without Lookahead Search
• Solving Texas Hold’em with Monte-Carlo Planning
• Reinforcement Learning of High-Speed Autonomous Driving through Unknown Map
• Implementation and deployment of particle filter for simulated and real-world localization tasks
• Tree Augmented Naive Bayes and Backward Simulation
• Transfer of Q values across tasks in Reinforcement Learning
• Training Regime Modifications for Deep Q-Network Learning Acceleration
• Reinforce Optimizer
• Approximating Ligand Docking Using a Markov Decision Process
• Breaking Down Social Media Filter Bubbles via Reinforcement Learning
• Performing an N-Sentiment Classification Task on Tinder Profiles Based On Image Feature Extraction
• Play Blackjack With Monte Carlo Simulation And Q-learning with Linear Regression
• Observer-Actor Neural Networks for Self-Play in Imperfect Information Games
• Using Hybrid Bayes Nets to Model Country Prosperity
• Solving a Pandemic! Various Approaches for Tackling the Board Game
• Improved Markov Decision Process Model for Resource Allocation in Disaster Scenarios
• Learning Chess through Reinforcement Learning
• Deep Reinforcement Learning For Continuous Control: An Investigation of Techniques and Tricks
• Computer Vision Through Perception: Semantic Understanding of Novel Scenes through Data Programming
• Path Planning for Insertion of a Bevel-Tip Needle
• Modeling human biases through reinforcement learning
• Bootstrapping Neural Network with Auxiliary Tasks
• Q-Learning Application in Optimizing Pokémon Battle Strategy
• Model-based exploration in natural language generation
• Automated Aircraft Touchdown
• Longitudinal Vehicle Control using a Markov Decision Process and Deep Neural Network
• MOMDP-based Aerial Target Search Optimization
• Greedy Thick-Thinning Structure Learning and Bayesian Network Conditional Independence Implementations in BayesNets.jl
• Multiagent Planning For Aerial Broadband Internet
• Viral Marketing as an MDP
• Neural Soccer – Towards Exploration by the Pursuit of Novelty
• Locally Weighted Value Iteration in Julia
• Fully-Nested Interactive POMDPs for Partially-Observable Turn-Based Games
• Optimal Policy Considerations for Gas Turbine Maintenance
• Learning Optimal Manipulation of Food Webs
• Estimating Resource Prospector’s Probability of Failure Using Importance Sampling and Cross Entropy
• Dynamically Discount Deep Reinforcement Learning
• Deep Reinforcement Learning: Accelerated Learning with Effective Gradient Ascent Optimization Algorithms
• Autonomous Human Tracking in Simulated Environment
• A LQG Library for POMDPs.jl

Fall 2015

Leduc Holdem Rules

• Mars Hab-Bot: Using MDPs to simulate a robot constructing human-livable habitats on Mars
• A Value Iteration Study of BlackJack
• Optimized Store-Stocking via Monte Carlo Tree Search with Stochastic Rewards
• Trajectory Planning for Map Exploration Using Terrain Features
• Instruction Following with Deep Reinforcement Learning
• Using Markov Decision Processes to Minimize Golf Score
• Reinforcement Learning for Scheduling I/O Requests on Multi-Queue Flash Storage
• Finding the Perfect ‘Job’ in resource allocation
• Maximizing Influence in Social Networks
• A Machine Learning Regression Approach to General Game Playing
• Modeling GPS Spoofing as a Partially Observable Markov Decision Process
• Travel Hacking with MDPs
• Optimal Mission Planning for a Satellite-Based Particle Detector via Online Reinforcement Learning
• An MDP Approach to Motion Planning for Hopping Rovers on Small Solar System Bodies
• Sampling Strategies for Deep Reinforcement Learning
• Descriptive Power of Bayesian Structure Learning in Stock Market
• Large-Scale Traffic Grid Signal Control Using Fuzzy Logic and Decentralized Reinforcement Learning
• Simulated Pedestrian-like Navigation with a 1D Kalman Filter with an Accelerometer and the Global Positioning System
• Search and Track Tradeoff for Multifunction Radars
• Play Calling in American Football Using Value Iteration
• Reinforcement learning for commodity trading
• Learning the Stock Market, a Naive Approach
• A POMDP Framework for Modelling Robotic Guidance During a Tissue Palpation Task
• Reinforcement Learning of an Artificially Intelligent Hearts Player1
• Toy Helicopter Control via Deep Reinforcement Learning
• Gas Refuelling Optimization Modelled as a Markov Decision Process
• Q-Matrix and Policy Compression via Deep Learning
• Augmenting Self-Learning In Chess Through Expert Imitation
• Monte Carlo Tree Search Applied to a Variant of the RockSample Problem
• Supply Chain Management using POMDPs
• Online Markov Decision Process Framework for Modeling Efficient Home Robot Cleaners
• Reinforcement Learning for Path Planning with Soft Robotic Manipulators
• Exploring POMDPS with Recurrent Neural Networks
• Tic-tac-toe with reinforcement learning: best strategies and influence of parameters
• Vehicle Speed Prediction using Long Short-Term Memory Networks
• Explorations on Learning Bayesian Networks
• Playing unknown game on a visual world
• Reinforcement Learning for Atari Games
• Q-learning in the Game of Mastermind
• Modeling of a Baseball Inning as MDP
• Reinforcement Learning for Path Planning with Soft Robotic Manipulators
• Autonomous Driving on a Multi-lane Highway Using POMDPs
• Solving a Maze Without Location Data
• Markov Decision Processes and Optimal Policy Determination for Street Parking
• Solving an opponent-based match-three mobile game
• Life begins as a POMDP: improving decision making in the IVF clinic
• Path Planning for Target-Tracking Unmanned Aerial Vehicle
• Discrete State Filter Implementation for a Battleships Artificial Intelligence
• POMDP for Search and Rescue with Obstacle Avoidance: Incorporation of Human in the Loop
• Application performance over cellular networks
• An MDP Approach to Motion Planning for Hopping Rovers on Small Solar System Bodies
• Solving Dudo: beating Liar’s Dice with a POMDP
• Reinforcement Learning for Tetris
• Robot Path Planning using Monte Carlo POMDP
• Reinforcement Learning of an Artificially Intelligent Hearts Player
• Enhancing Computational Efficiency of PILCO Model-based Reinforcement Learning Algorithm
• Analysis of UCT Exploration Parameter in Sailing Domain Problems
• Solving a Search and Rescue Planning problem with MOMDPs
• Robot Motion Planning in Unknown Environments using Monte Carlo Tree Search
• Delivery optimization of an on-demand delivery service
• Solving Multi­Agent Decision Making using MDPs
• Efficient and Modular Inventory Management Framework for Small Businesseses
• Markov Decision Processes in Board Game Playing
• Automated Model Selection via Gaussian Processes
• Predictive Hybrid Vehicle Control Policy
• Optimal Policies for In-Space Satellite Communications
• Spacecraft Navigation in Cluttered, Dynamic Environments Using 3D Lidar
• Playing Chess Endgames using Reinforcement Learning
• Space Debris Removal
• Large-Scale Traffic Grid Signal Control Using Fuzzy Logic and Decentralized Reinforcement Learning
• Relation Extraction from Scratch
• Lane Merging as a Markov Decision Process
• Using MDP/POMDP to Help in Search of Survivors of a Plane Crash
• Applying POMCP to Controlling Partially Observable Diffusion Processes
• Credit Risk Classification using Bayesian Network

Leduc Hold'em Tournaments

Fall 2014

Leduc Hold'em

• Automating Air Traffic Management for Flight Arrivals
• Policy Learning for Sokoban
• Flight Path Optimization Under Constraints Using a Markov Decision Process Approach
• Visual Localization and POMDP for Autonomous Indoor Navigation
• Monte Carlo Tree Search for Online Learning in Golf Course Management
• Pushing on Leaves
• Beating 2048
• Improved electrical grid balancing with demand response scheduled by an MDP
• Multi-Fidelity Model Management in Engineering Design Optimization Using Partially Observable Markov Decision Processes
• Smarter Generators in Power Markets
• Beach Paddle Ball
• Applying POMDP to RockSample problem
• Targeting Hostile Vehicle Modeled as a Partially Observable Markov Decision Process with State-Dependent Observation Model
• Reinforcement Learning and Linear Gaussian Dynamics Applied to Multifidelity Optimization of a Supersonic Wedge
• Approximate POMDP Solutions for Short-Range UAV Traffic Conflict Resolution
• WorkSmart: The Implementation of a Modified Q-Learning Algorithm for an Intelligent Daily To-Do List Android Application
• Imminent Obstacle Avoidance with Friction Uncertainty
• Dynamic Restrictions during Commercial Space Vehicle Launches
• Autonomous Direct Marketing with Deep Q-Learning
• Efficient Risk Estimation for Chance-Constrained Robotic Motion Planning Under Uncertainty
• Probabilistic Aircraft Arrival Rate Prediction
• Audio Keylogging: Translating Acoustic Signals into Keystrokes
• Collision Avoidance for Small Multi-Rotor Aircraft using SARSA(Phone Slots For Real Money
• Games With Real Money
• Online Sports Betting In Tennessee
• Aristocrat Online Casino
• Online Casino Ipad