For nearly two decades, Kenneth Forbus and Johan de Kleer have accumulated a substantial body of knowledge about the principles and practice of creating problem solvers. In some cases they are the inventors of the ideas or techniques described, and in others, participants in their development. Building Problem Solvers communicates this knowledge in a focused, cohesive manner. It is unique among standard artificial intelligence texts in combining science and engineering, theory and craft to describe the construction of AI reasoning systems, and it includes code illustrating the ideas. After working through Building Problem Solvers, readers should have a deep understanding of pattern directed inference systems, constraint languages, and truth-maintenance systems.

We consider the problem of generating optimal stochastic policies for Constrained Stochastic Shortest Path problems, which are a natural model for planning under uncertainty for resource-bounded agents with multiple competing objectives. While unconstrained SSPs enjoy a multitude of efficient heuristic search solution methods with the ability to focus on promising areas reachable from the initial state, the state of the art for constrained SSPs revolves around linear and dynamic programming algorithms which explore the entire state space. In this paper, we present i-dual, which, to the best of our knowledge, is the first heuristic search algorithm for constrained SSPs. To concisely represent constraints and efficiently decide their violation, i-dual operates in the space of dual variables describing the policy occupation measures. It does so while retaining the ability to use standard value function heuristics computed by well-known methods. Our experiments on a suite of PPDDL problems augmented with constraints show that these features enable i-dual to achieve up to two orders of magnitude improvement in run-time and memory over linear programming algorithms.

This paper addresses the challenge of automated numeric domain model acquisition from observations. Many industrial and commercial applications of planning technology rely on numeric planning models. For example, in the area of autonomous systems and robotics, an autonomous robot often has to reason about its position in space, power levels and storage capacities. It is essential for these models to be easy to construct. Ideally, they should be automatically constructed. Learning the structure of planning domains from observations of action traces has produced successful results in classical planning. In this work, we present the first results in generalising approaches from classical planning to numeric planning. We restrict the numeric domains to those that include fixed action costs. Taking the finite state automata generated by the LOCM family of algorithms, we learn costs associated with machines; specifically to the object transitions and the state parameters. We learn action costs from action traces (with only the final cost of the plans as extra information) using a constraint programming approach. We demonstrate the effectiveness of this approach on standard benchmarks.

The program works in two phases: In the first phase it detects the permutation of the cube. The faces of the cube can be shown in arbitrary order. Then, as soon as it is confident about the permutation of the cube, it computes how to solve it in 20 turns or less and shows the first turn directly on the cube. It recognizes whenever a turn is done and shows the next one right away. There is no other input from the user other than the data from the camera.

This report is my final project for the MIT Media Lab Class "Integrative Theories of Mind and Cognition" (also known as Future of AI, and New Destinations in Artificial Intelligence) in Spring 2016. Artificial Intelligence performs gradient descent. The AI field discovers a path of success, and then travels that path until progress stops (when a local minimum is reached). Then, the field resets and chooses a new path, thus repeating the process. If this trend continues, AI should soon reach a local minimum, causing the next AI winter. However, recent methods provide an opportunity to escape the local minimum. To continue recent success, it is necessary to compare the current progress to all prior progress in AI. I begin this paper by pointing out a concerning pattern in the field of AI and describing how it can be useful to model the field's behavior. The paper is then divided into two main sections. In the first section of this paper, I argue that the field of artificial intelligence, itself, has been performing gradient descent. I catalog a repeating trend in the field: a string of successes, followed by a sudden crash, followed by a change in direction. In the second section, I describe steps that should be taken to prevent the current trends from falling into a local minimum. I present a number of examples from the past that deep learning techniques are currently unable to accomplish. Finally, I summarize my findings and conclude by reiterating the use of the gradient descent model.

All chapters are available to download as of Sept 2015.

If yesterday's Google I/O keynote was anything to go by, it's abundantly clear that Artificial Intelligence is here to stay. Many of the mobile apps and services we use each day already have some level of AI powering their internals, but these days it's become so seamless that we hardly ever notice. I frankly don't talk to my phone too often, but when I do--from asking Google Now to ask for information about the new city I've landed in, to querying Google Photos to show me all photos of my daughter laughing, it's fascinating (and sometimes scary) to see these tools in surprisingly accurate action. Many of these advancements are thanks to rapid developments in the fields of machine learning and artificial intelligence, developments that now enable computing devices to understand our voice, or recognize features in photos. Coupled with cloud-based services and cellular connectivity, all of this'intelligence' is suddenly available to billions of smartphones, from mega-metropolises to humble villages.

In a case proposed by the research team, a device like a flat smartphone could be folded and reconfigured into the shape of a game controller. Or, in a less practical example, you could simply roll your phone out into a rectangular log with a postage-stamp sized display on one end. For users who were never very good at spatial reasoning or origami, an algorithm will help determine the best way to twist and fold the screen into the desired shape. While the device is still in the awkward prototype phase at this point, the research team will present it to a panel at the International Conference on Robotics and Automation in Stockholm later this week.

Commonsense knowledge representation and reasoning is key for tasks such as artificial intelligence and natural language understanding. Since commonsense consists of information that humans take for granted, gathering it is an extremely difficult task. In this paper, we introduce a novel 3D game engine for commonsense knowledge acquisition (GECKA3D) which aims to collect commonsense from game designers through the development of serious games. GECKA3D integrates the potential of serious games and games with a purpose. This provides a platform for the acquisition of re-usable and multi-purpose knowledge, and also enables the development of games that can provide entertainment value and teach players something meaningful about the actual world they live in.

The concept of "affordance" represents the relationship between human perceivers and their environment. Affordance perception, representation, and inference are central to commonsense reasoning, tool-use and creative problem-solving in artificial agents. Existing approaches fail to provide flexibility with which to reason about affordances in the open world, where they are influenced by changing context, social norms, historical precedence, and uncertainty. We develop a formal rules-based logical representational format coupled with an uncertainty-processing framework to reason about cognitive affordances in a more general manner than shown in the existing literature. Our framework allows agents to make deductive and abductive inferences about functional and social affordances, collectively and dynamically, thereby allowing the agent to adapt to changing conditions. We demonstrate our approach with an example, and show that an agent can successfully reason through situations that involve a tight interplay between various social and functional norms.