In order to cooperate effectively with human players, characters need to infer the tasks players are pursuing and select contextually appropriate responses. This process of parsing a serial input stream of observations to infer a hierarchical task structure is much like the process of compiling source code. We draw an analogy between compiling source code and compiling behavior, and propose modeling the cognitive system of a character as a compiler, which tokenizes observations and infers a hierarchical task structure. An evaluation comparing automatically compiled behavior to human annotation demonstrates the potential for this approach to enable AI characters to understand the behavior and infer the tasks of human partners.

Learning effective policies in multi-agent adversarial games is a significant challenge since the search space can be prohibitively large when the actions of all the agents are considered simultaneously. Recent advances in Monte Carlo search methods have produced good results in single-agent games like Go with very large search spaces. In this paper, we propose a variation on the Monte Carlo method, UCT (Upper Confidence Bound Trees), for multi-agent, continuous-valued, adversarial games and demonstrate its utility at generating American football plays for Rush Football 2008. In football, like in many other multi-agent games, the actions of all of the agents are not equally crucial to gameplay success. By automatically identifying key players from historical game play, we can focus the UCT search on player groupings that have the largest impact on yardage gains in a particular formation.

In this paper, we report progress on making synthetic characters more taskable. In particular, we present an English-like command language that lets one specify complex behaviors an agent should carry out in a virtual environment. We also report compilers that translate English commands into a formal notation and formal statements into procedures for Icarus, an agent architecture that supports reactive execution. To demonstrate the benefits of such taskability, we have integrated Icarus with Twig, which provides a simulated physical environment with humanoid agents. We use the command language to specify three complex activities, including responding to an object contingently, collecting and storing a set of objects, and negotiating with another agent in order to purchase an item. We also discuss related work on controlling synthetic characters, along with paths for additional research on taskability.

Players begin games at different skill levels and develop their skill at different rates—so that even the best-designed games are uninterestingly easy for some players and frustratingly difficult for others. A proposed answer to this challenge is Dynamic Difficulty Adjustment (DDA), a general category of approaches that alter games during play, in response to player performance. However, nearly all these techniques are focused on basic parameter tweaking, while the difficulty of many games is connected to aspects that are more challenging to adjust dynamically, such as level design. Further, most DDA techniques are based on designer intuition, which may not reflect actual play patterns. Responding to these challenges, we have created Polymorph, which employs techniques from level generation and machine learning to understand level difficulty and player skill, dynamically constructing levels for a 2D platformer game with continually-appropriate challenge. We present the results of the user study on which Polymorph's model of level difficulty is based, as well as a discussion of the unique features of the model. We believe Polymorph creates a play experience that is unique because the changes are both personalized and structural, while also providing an example of a new application of machine learning to aid game design.

Team coordination of non-player characters can create a deeper sense of immersion in real-time games by allowing characters to work together to produce better tactics and strategy. Achieving multi-agent coordination can be a difficult problem, and can incur substantial computational costs. Our goal with this work is to produce a reactive method for coordinating game characters that will allow computationally inexpensive team coordination. Reactive teaming creates teams of agents through the use of simple constant-time agent interactions without increasing the difficulty of authoring game characters.

We have developed and demonstrated an experimental authoring and run-time tool, called Disco for Games, that supports the creation of games in which dialogue and action are integrated without the need for changing modes. This tool is based on collaborative discourse theory and hierarchical task networks, in which utterances are treated as actions, and has a number of additional benefits including better modeling of interruptions, automatic dialogue generation, plan recognition and automatic failure retry.

Game environments are becoming more and more mutable from the actions of both Players and Non Player Characters (NPCs). However, current generation AI agents do not take advantage of the tactical abilities these mutable worlds provide. We propose a method to make the game agents aware of the mutability of the world by extending their repertoire of abilities to include world alteration commands and some evaluation functions, which determine when and where to alter the world for the greatest tactical gain. Primarily, our work focuses on the Adversarial Navigation Mesh Alteration (ANMA) algorithm, which evaluates potential changes to the map in adversarial environments from an attacker and defender point of view. We present an empirical evaluation of the ANMA algorithm in a Capture The Flag (CTF) simulation environment with several teams of agents. One group of agents (adaptive) lacks the ability to initiate world deformations, but they can respond and re-plan to take advantage of world modifications. The second team of agents (builders) can only generate additional paths through the world using the attacker portion of ANMA. The third team of agents (universal) is able to fully deform the world by generating new paths or removing existing paths using both the attacker and defender sections of ANMA. We evaluated these teams and observed that builder agents beat adaptive agents at a rate of 1.33 to 1. The more advanced universal agents beat adaptive agents at a rate of 2.75 to 1 and builder agents 1.4 to 1.

Massively Multiplayer Online Games (MMOGs), in their aspect as online communities, represent an exciting opportunity for studying social and behavioral models.  For that purpose we have developed Cosmopolis, a free MMOG containing several key research-oriented features.  First, Cosmopolis consists of an outer game for larger-scale social modeling, as well as a set of subgames suitable for tightly-controlled sandbox-style experiments, all allowing a high level of data logging configuration and control by researchers.  Also, Cosmopolis’s world model incorporates configurable, AI-driven non-player character communities, as a means of researching interactions between individuals and societies

One of the main challenges in game AI is building agents that can intelligently react to unforeseen game situations. In real-time strategy games, players create new strategies and tactics that were not anticipated during development. In order to build agents capable of adapting to these types of events, we advocate the development of agents that reason about their goals in response to unanticipated game events. This results in a decoupling between the goal selection and goal execution logic in an agent. We present a reactive planning implementation of the Goal-Driven Autonomy conceptual model and demonstrate its application in StarCraft. Our system achieves a win rate of 73% against the built-in AI and outranks 48% of human players on a competitive ladder server.

Procedural content generation is becoming more and more relevant to solve the problem of content creation for the ever growing virtual worlds of games, simulations and other applications. However, these procedures are often unintuitive or use vague parameters, making it somewhat difficult for a designer to express his or her creative intent. Even worse, most of these techniques lack an accessible and easy to use interface.We have developed a generic layout solving approach to automatically create sensible content for virtual worlds. In that context, this paper proposes a high-level scene description language that allows designers to specify particular types of scenes. This description language allows designers to easily specify which objects need to be present in a scene, their attributes, and possible interrelationships. Application of the language, based on the rich vocabulary taken from a semantic library, is illustrated with several examples, showing its flexibility, intuitiveness and ease of use.