This paper presents a method for imbuing agents with high-level domain knowledge via short command phrases given by a human designer. These command phrases, or goals, are provided at design time and describe the behaviors a human deems necessary in order to succeed at the current task. At run time, each goal is converted into a vector representation and is used to guide the agent’s exploration in a text-based environment. At each time step, the agent selects behaviors which it believes are relevant to its current objective, thus reducing the search space and increasing obtained rewards. We show that task-specific goals improve performance in a majority of virtual worlds, and discuss the potential for more general applications.

In the world of online gaming, not all actions are created equal. For example, when a player’s character is confronted with a closed door, it would not make much sense to brandish a weapon, apply a healing potion, or attempt to barter. A more reasonable response would be to either open or unlock the door. The term interaction mode embodies the idea that many potential actions are neither useful nor applicable in a given situation. This paper presents a AEGIM, an algorithm for the automated extraction of game interaction modes via a semantic embedding space. AEGIM uses an image captioning system in conjunction with a semantic vector space model to create a gestalt representation of in-game screenshots, thus enabling it to detect the interaction mode evoked by the game.

We present BYU-EVE, an open domain dialogue architecture that combines the strengths of hand-crafted rules, deep learning, and structured knowledge graph traversal in order to create satisfying user experiences. Rather than viewing dialogue as a strict mapping between input and output texts, EVE treats conversations as a collaborative process in which two jointly coordinating agents chart a trajectory through experiential space. A key element of this architecture is the use of conversational scaffolding, a technique which uses a (small) conversational dataset to define a generalized response strategy. We also take the innovative approach of integrating the agent’s self and user models directly within the knowledge graph. This allows EVE to discern topics of shared interest while simultaneously identifying areas of ambiguity or cognitive dissonance.

Autonomous agents require a grounded understanding of language, and multimodal embeddings seem like a logical way to accomplish this. But once you have a joint embedding space, what is it good for? How do you use it? And most especially, how do you map human input (such as verbal instructions) into a set of behaviors that conform with the human’s desires? This paper presents a method for using distributional sentence representations to prioritize action primitives. Inspired by the human ability to transfer domain knowledge via spoken language, we provide our agent with natural language instructions that describe high-reward behaviors. The agent then uses these instructions to align its actions with human intents. A key advantage of this approach is its flexibility: once a properly-structured embedding space has been obtained, new mappings between instructions and behaviors do not have to be trained using thousands or tens of thousands of labeled examples. Less than twenty data points is sufficient.

Autonomous agents must often detect affordances: the set of behaviors enabled by a situation. Affordance detection is particularly helpful in domains with large action spaces, allowing the agent to prune its search space by avoiding futile behaviors. This paper presents a method for affordance extraction via word embeddings trained on a Wikipedia corpus. The resulting word vectors are treated as a common knowledge database which can be queried using linear algebra. We apply this method to a reinforcement learning agent in a text-only environment and show that affordance-based action selection improves performance most of the time. Our method increases the computational complexity of each learning step but significantly reduces the total number of steps needed. In addition, the agent's action selections begin to resemble those a human would choose.

Autonomous agents must often detect affordances: the set of behaviors enabled by a situation. Affordance detection is particularly helpful in domains with large action spaces, allowing the agent to prune its search space by avoiding futile behaviors. This paper presents a method for affordance extraction via word embeddings trained on a tagged Wikipedia corpus. The resulting word vectors are treated as a common knowledge database which can be queried using linear algebra. We apply this method to a reinforcement learning agent in a text-only environment and show that affordance-based action selection improves performance in most cases. Our method increases the computational complexity of each learning step but significantly reduces the total number of steps needed. In addition, the agent’s action selections begin to resemble those a human would choose.

Simultaneous multi-user computer aided design (CAD) allows multiple designers to contribute to the same model at the same time. The resulting parallel design workflow shortens product development cycles. In a replicated, simultaneous multi-user CAD system, modeling data must be kept consistent between clients. This paper presents a method that keeps independent copies of the models in sync between distributed CAD clients. This is accomplished by enforcing modeling operations to occur in the same order on all the clients. In case of conflict, a resolution method preserves conflicting operations locally for later reuse or resolution by the user. These methods are implemented in a commercial CAD system which has been enhanced to enable simultaneous multi-user. Validation tests are run to demonstrate that the methods implemented ensure model consistency and resolve conflicts while preserving conflicting operation data.