Legged robots have consistently captured our collective imagination through various forms of media, from Hollywood films, anime, and viral Youtube videos of robots accomplishing incredible feats of acrobatics. These robots have the potential to navigate our environments, capable of completing tasks that would otherwise require human intervention. However, developing controls...
In recent years, model-free Deep Reinforcement Learning (RL) has become an increasingly popular alternative to more traditional model-based or optimization-based control methods in solving robotic legged locomotion. However, deploying RL in the real world can be a significant undertaking. Constructing reward functions which compel controllers to learn the desired behavior...
Learning latent space representations of high-dimensional world states has been at the core of recent rapid growth in reinforcement learning(RL). At the same time, RL algo- rithms have suffered from ignored uncertainties in the predicted estimates of model-free or model-based methods. In our work, we investigate both of these aspects...
Reinforcement learning has emerged as a popular tool for solving control tasks, with multiple works focusing on the complex and dynamic task of locomotion. However, the naive application of reinforcement learning to this problem often produces maladaptive policies that exploit the model or reward function. This results in behavior that...
Papers proposing novel machine learning algorithms tend to present the algorithm or technique in question in the best possible light. The standard practice is generally for authors to emphasize their proposed algorithms' performance in the precise setting where it is maximally impressive, often by only fully evaluating their best known...
We take for granted how quickly we, as humans, form mental models of the world around us. By the time we are toddlers, we have an observable intuition around the physical rules of the world. Stacking blocks such that they don’t fall over becomes such a trivial task, that it...
Dynamic bipedal locomotion is among the most difficult and yet relevant problems in modern robotics. While a multitude of classical control methods for bipedal locomotion exist, they are often brittle or limited in capability. In recent years, work in applying reinforcement learning to robotics has lead to superior performance across...
Anomaly detection has been used in variety of applications in practice, including cyber-security, fraud detection and detecting faults in safety critical systems, etc. Anomaly detectors produce a ranked list of statistical anomalies, which are typically examined by human analysts in order to extract the actual anomalies of interest. Unfortunately, most...
This thesis considers the problem of training convolutional neural networks for online visual tracking. A major challenge for single object visual tracking is that most training sets with frame-level track annotations are quite small, due to the prohibitive cost of manual annotation. Current training approaches either supplement the annotations with...
This thesis focuses on the problem of object tracking. Given a video, the general objective of tracking is to track the location over time of one or more targets in the image sequence. This is a very challenging task as algorithms need to deal with problems such as appearance variations,...
We model the popular board game of Clue as an MDP and evaluate Monte-Carlo policy rollout in a simulated environment pitting different agents and policies against each other. We describe the choices we made in the representation, along with some of the problems we encountered along the way. We find...
Monte-Carlo Tree Search (MCTS) is an online-planning algorithm for decision-theoretic planning in domains with stochastic and combinatorial structure. The general applicability of MCTS makes it an ideal first choice to investigate when developing planners for complex applications requiring automated control and planning. The first contribution of this thesis is to...
Automatic analysis of American football videos can help teams develop strategies and extract patterns with less human effort. In this work, we focus on the problem of automatically determining which team is on offense/defense, which is an important subproblem for higher-level analysis. While seemingly mundane, this problem is quite challenging...
We are witnessing the rise of the data-driven science paradigm, in which massive amounts of data - much of it collected as a side-effect of ordinary human activity - can be analyzed to make sense of the data and to make useful predictions. To fully realize the promise of this...
We investigate a search and coverage planning problem, where an area of interest has to be explored by a number of vehicles, given a fixed time budget. A good coverage plan has a low probability of a target remaining unobserved. We introduce a formal problem statement, suggest a greedy algorithm...
A five year study of file-system metadata shows that the number of files increases by 200% and only a select few file-types contribute for over 35% of the files that exist on a file-system. It is difficult to point out a permanent selection of files that a user really cares...
This paper describes a method for finding areas of interest on a two-dimensional grid map used in the real-time strategy engine Stratagus. The method involves discovering chokepoints where through all simulation agents must pass. Using a set of tunable parameters, a full set of chokepoints are located. The redundant and...
How can an agent generalize its knowledge to new circumstances? To learn
effectively an agent acting in a sequential decision problem must make intelligent action selection choices based on its available knowledge. This dissertation focuses on Bayesian methods of representing learned knowledge and develops novel algorithms that exploit the represented...
Efficient routing of information packets in dynamically changing communication networks requires routing policies that adapt to changes in load levels, traffic patterns and network topologies. Reinforcement Learning (RL) is an area of artificial intelligence that studies algorithms that dynamically optimize their performance based on experience in an environment. RL, thus,...
Generating solutions to Sokoban levels is an NP-hard problem that is difficult for even modern day computers to solve due to its complexity. This project explores the creation of a Sokoban solver by eliminating as many potential moves as possible to greatly limit the overall search space. This reduction is...