In this paper, we present a formalism we call Box-Split Grammars for the procedural modeling of structures in Minecraft and similar environments. Our grammars are based on previous work on split grammars and box grammars, where rules define how a given box, labeled with a non-terminal symbol, can be split into smaller pieces, and how subsequent rules are to be applied. We represent grammar rules as ordinary, well-structured python functions, allowing the integration into existing systems, and demonstrate their utility by recreating variations of ancient Greek temples using a few simple grammar rules.

Cooperation between AI agents and humans is of ever greater importance. In this paper we present an AI agent for the game Pandemic that was specifically designed to play cooperatively with a human player. Our agent utilizes planning to determine which actions to perform, and plan recognition to determine the current goal of its cooperator in order to assist them. We also present an experiment we performed with human participants, and how our agent performs at a level that is comparable to other AI agents playing with themselves, when playing with a human player, as well as the impact of plan recognition on how the participants perceive the AI agent.

The cooperative card gameHanabihas become of increasing interest in the community, since it combines partially hidden information with information exchange using restricted communication channels. In this article, we describe artificial intelligence agents that are designed to play the game with human players. Our agents make use of the fact that human players expect other players to act intentionally by formulating goals of their own and planning how to achieve them. They then use the available actions available to communicate their plan to the human player. On the flip side, our agents also interpret the actions performed by the human player as containing information about their plans. We present two different variants of our agent that perform this interpretation in different ways. Additionally, since part of human communication happens in subtle indirect ways, we also demonstrate that our agent can use the timing of the human player’s actions as additional information. In order to validate our agents, we have performed two separate experiments: one was done to validate the intentional component of the agents, while the other focused on the interpretation of received information. In this article, we also present the results obtained from these two experiments.

Hearthstone is a competitive, online Collectible Card Game, in which players construct their own 30-card decks from hundreds of available cards. Different decks differ wildly in terms of their strategy, from very agressive decks that seek to attack the opponent early, to decks relying on certain combinations of cards, to decks that are focused on responding to the opponent’s and ending the game slowly. The player community has therefore given names to different deck archetypes, depending on the strategy they pursue. When playing the game, knowing which archetype the opponent’s deck is likely to have helps inform a player on how they should adapt their own strategy to best counter the opponent’s. In this paper we introduce the problem of predicting a player’s deck archetype from minimal information, in particular only from the actions they performed on their first turn. We discuss the relevance of this problem, and how it can help players adapt to the opponent’s strategy, as well as information that can be learned from it. While the information was intentionally chosen to be minimal, due to the nature of the game it still varies in size from game to game, which presents an additional challenge. We describe different approaches to handle this information and their performance applied to this problem, comparing standard statistical methods with Recurrent Neural Networks, and their relative trade-offs, in particular with regards to training time.

Hanabi is a cooperative card game in which communication plays a key role. The game provides an interesting challenge for AI agents, because the game state is only partially observable, and the game limits what players can tell each other. This limit on communication channels is similar to a common scenario in system security research, and has been researched extensively in that context, for example by bypassing a system's isolation by establishing a covert communication channel. Such channels can be established through anything that the sending party can influence and the receiving party can observe, such as photonic emission, resource contention, or latency. In this paper, we present Hanabi agents that utilize timing as a covert channel so effectively that they can eschew the communicative actions provided by the game entirely. In addition to a thorough evaluation of the effectiveness of our approach, and a comparison to other Hanabi agents, we provide its context in the area of security, and an outlook on how it could be related to human behavior in future work.