Probabilistic models were fit to logs of player actions in the card game Dominion in an attempt to find evidence of personality types that could be used to classify player behavior as well as generate probabilistic bot behavior. Expectation Maximization seeded with players' self-assessments for their motivations was run for two different model types — Naive Bayes and a trigram model — to uncover three clusters each. For both model structures, most players were classified as belonging to a single large cluster that combined the goals of splashy plays, clever combos, and effective play, cross-cutting the original categories — a cautionary tale for research that assumes players can be classified into one category or another. However, subjects qualitatively report that the different model structures play very differently, with the Naive Bayes model more creatively combining cards.

Hiding and seeking is a cognitive ability frequently demonstrated by humans in both real life and video games. We use machine learning to automatically construct the first computational model of hide/seek behavior in adult humans in a video game like setting. The model is then run generatively in a novel environment and its behavior is found indistinguishable from actual human behavior by a panel of human judges.  In doing so the artificial intelligence agent using the model appears to have passed a version of the Turing test for hiding and seeking.

In this paper, we apply the idea of testing games by learning interactions with them that cause unwanted behavior of the game to test the competition entries for some of the scenarios of the 2010 StarCraft AI competition. By extending the previously published macro action concept to include macro action sequences for individual game units, by adjusting the concept to the real-time requirements of StarCraft, and by using macros involving specific abilities of game units, our testing system was able to find either weaknesses or system crashes for all of the competition entries of the chosen scenarios. Additionally, by requiring a minimal margin with respect to surviving units, we were able to clearly identify the weaknesses of the tested AIs.

Massive Multiplayer Online Strategy games present several unique challenges to players and designers. There is the need to constantly adapt to changes in the game itself and the need to achieve a certain level of simulation and realism, which typically implies battles involving combat with several distinct armies, combat phases and diferent terrains; resource management which involves buying and selling goods and combining lots of diferent kinds of resources to fund the player's nation and cutthroat diplomacy which dictates the pace of the game. However, these constant changes and simulation mechanisms make a game harder to play, increasing the amount of effort required to play it properly. As some of these games take months to be played, players who become inactive have a negative impact on the game. This work pretends to demonstrate how to create versatile agents for playing Massive Multiplayer Online Turn Based Strategy Games, while keeping close attention to their playing performance. In a test to measure this performance the results showed similar survival performance between humans and AIs.

A big challenge for creating human-level game AI is building agents capable of operating in imperfect information environments. In real-time strategy games the technological progress of an opponent and locations of enemy units are partially observable. To overcome this limitation, we explore a particle-based approach for estimating the location of enemy units that have been encountered. We represent state estimation as an optimization problem, and automatically learn parameters for the particle model by mining a corpus of expert StarCraft replays. The particle model tracks opponent units and provides conditions for activating tactical behaviors in our StarCraft bot. Our results show that incorporating a learned particle model improves the performance of EISBot by 10% over baseline approaches.

The creation of effective autonomous agents (bots) for combat scenarios has long been a goal of the gaming industry. However, a secondary consideration is whether the autonomous bots behave like human players; this is especially important for simulation/training applications which aim to instruct participants in real-world tasks. Bots often compensate for a lack of combat acumen with advantages such as accurate targeting, predefined navigational networks, and perfect world knowledge, which makes them challenging but often predictable opponents. In this paper, we examine the problem of teaching a bot to play like a human in first-person shooter game combat scenarios. Our bot learns attack, exploration and targeting policies from data collected from expert human player demonstrations in Unreal Tournament. We hypothesize that one key difference between human players and autonomous bots lies in the relative valuation of game states. To capture the internal model used by expert human players to evaluate the benefits of different actions, we use inverse reinforcement learning to learn rewards for different game states. We report the results of a human subjects' study evaluating the performance of bot policies learned from human demonstration against a set of standard bot policies. Our study reveals that human players found our bots to be significantly more human-like than the standard bots during play. Our technique represents a promising stepping-stone toward addressing challenges such as the Bot Turing Test (the CIG Bot 2K Competition).

Grid representations offer many advantages for path planning. Lookups in grids are fast, due to the uniform memory layout, and it is easy to modify grids. But, grids often have significant memory requirements, they cannot directly represent more complex surfaces, and path planning is slower due to their high granularity representation of the world. The speed of path planning on grids has been addressed using abstract representations, such as has been documented in work on Dragon Age: Origins. The abstract representation used in this game was compact, preventing permanent changes to the grid. In this paper we introduce a sparse grid representation, where grid cells are only stored where necessary. From this sparse representation we incrementally build an abstract graph which represents possible movement in the world at a high-level of granularity. This sparse representation also allows the representation of three-dimensional worlds. This representation allows the world to be incrementally changed in under a millisecond, reducing the maximum memory required to store a map and abstraction from Dragon Age: Origins by nearly one megabyte. Fundamentally, the representation allows previously allocated but unused memory to be used in ways that result in higher-quality planning and more intelligent agents.

We apply decision theoretic techniques to construct non-player characters that are able to assist a human player in collaborative games. The method is based on solving Markov decision processes, which can be difficult when the game state is described by many variables. To scale to more complex games, the method allows decomposition of a game task into subtasks, each of which can be modelled by a Markov decision process. Intention recognition is used to infer the subtask that the human is currently performing, allowing the helper to assist the human in performing the correct task. Experiments show that the method can be effective, giving near-human level performance in helping a human in a collaborative game.

This paper describes the creation of a digital improvisational theatre game, called Party Quirks, that allows a human user to improvise a scene with synthetic actors according to the rules of the real-world Party Quirks improv game. The AI actor behaviors are based on our study of communication strategies between real-life actors on stage and the fuzzy concepts that they employ to define and portray characters. This paper describes the underlying fuzzy concepts used to enable reasoning in ambiguous environments, like improv theatre. It also details the development of content for the system, which involved the creation of a system for animation authoring, design for efficient data reuse, and a work flow centered on Google Docs enabling parallel data entry and rapid iteration.

Many forms of interactive digital entertainment involve interacting with virtual dramatic characters. Our long term goal is to procedurally generate character dialogue behavior that automatically mimics, or blends, the style of existing characters. In this paper, we show how linguistic elements in character dialogue can define the style of characters in our RPG SpyFeet. We utilize a corpus of 862 film scripts from the IMSDb website, representing 7,400 characters, 664,000 lines of dialogue and 9,599,000 word tokens. We utilize counts of linguistic reflexes that have been used previously for personality or author recognition to discriminate different character types. With classification experiments, we show that different types of characters can be distinguished at accuracies up to 83% over a baseline of 20%. We discuss the characteristics of the learned models and show how they can be used to mimic particular film characters.