Multi-agent systems are used in domains where individual component autonomy and cooperation are necessary. The overall system performance requires that the diverse agents maintain quality interactions to facilitate cooperation. A complication to inter-agent interaction occurs when the agents learn (change their own functionality), when new agents are introduced, or existing agents are functionally modified. This research focuses on creating a general use multi-agent system, Middleware Unifying Framework for Independent Nodes System (MUFFINS), and implementing a mechanism, the Megagent, that addresses the interaction challenges. The Megagent provides the ability for agents to assess their performance per data source and to improve it with transformations based on feedback. Evaluation of the concept is tested on data mangled from the Digits dataset to represent learning and new agents and in all cases improves accuracy over a static agent.

Control architectures and algorithms for large autonomous swarms are receiving increased research interest. Control of swarm systems becomes more difficult as the size of the agent swarm increases, making centralized control approaches inadequate. This paper presents the informal team assignment algorithm. By leveraging agent roles and signaling actions, the algorithm provides a local agent mechanism leading to the emergence of cooperative teams. Tested in a modified pursuit-evasion domain, simulation results demonstrate that agent roles and inter-agent signaling spontaneously create small collaborative agent teams dedicated to shared task accomplishment. The result is in higher win ratios for signal and role capable swarms.

An increasingly important job for the autonomous agents is determining what goal they should be accomplishing. In dynamic environments the goal of the autonomous agents does not always remain constant. This research examines how to detect and adapt to goal changes within a dynamic game environment. An adaptive learner capable of detecting concept drift is used to detect when a goal change has occurred within the game environment and exploration techniques are used to adapt to the change. Initial results show that the agent has an 84% detection rate.

Many organizations with an in-house information technology department rely on a trouble ticket system to track network issues. The goal of an effective trouble ticket system is to prioritize limited support personnel, responsively address each issue, and maintain user satisfaction. This paper presents a machine learning system that predicts ticket resolution time to provide users with an expected resolution time for their issue upon ticket submission. Classification and regression models were developed using boosted regression trees and artificial neural networks (ANNs). Evaluating on 12,303 trouble tickets, the classification model accuracy from the boosted regression tree was 74.5%. As a regression problem, the ANN model achieved the best result, with a mean absolute error (MAE) of 24.8 hours.

Reinforcement learning (RL) is a particular machine learning technique enabling an agent to learn while interacting with its environment. Agents in non-stationary environments are faced with the additional problem of handling concept drift, which is a partially-observable change that modifies the environment without notification. This causes several problems: agents with a decaying exploration fail to adapt while agents capable of adapting may over fit to noise and overwrites previously learned knowledge. These issues are known as the plasticity-stability dilemma and catastrophic forgetting, respectively. Agents in such environments must take steps to mitigate both problems. This work contributes an algorithm that combines a concept drift classifier with multi-objective reinforcement learning (MORL) to produce an unsupervised technique for learning in non-stationary environments, especially in the face of partially observable changes. The algorithm manages the plasticity-stability dilemma by strategically adjusting learning rates and mitigates catastrophic forgetting by systematically storing knowledge and recalling it when it recognizes repeat situations. Results demonstrate that agents using this algorithm outperform agents using an approach that ignores non-stationarity.

The Model AI Assignments session seeks to gather and disseminate the best assignment designs of the Artificial Intelligence (AI) Education community. Recognizing that assignments form the core of student learning experience, we here present abstracts of six AI assignments from the 2017 session that are easily adoptable, playfully engaging, and flexible for a variety of instructor needs.

In this paper we examine the build order problem in real-time strategy (RTS) games in which the objective is to optimize execution of a strategy by scheduling actions with respect to a set of subgoals. We model the build order problem as a multi-objective problem (MOP), and solutions are generated utilizing a multi-objective evolutionary algorithm (MOEA). A three dimensional solution space is presented providing a depiction of a Pareto front for the build order MOP. Results of the online strategic planning tool are provided which demonstrate that our planner out-performs an expert scripted player. This is demonstrated for an AI agent in the Spring Engine Balanced Annihilation RTS game.

One of the defining characteristics of an adaptive automation system is the hand-off from machine to human--and vice versa. This research seeks to improve system control hand-offs, by investigating how the manner in which the automation completes its task affects the overall performance of the human-machine team. Specifically, the research will explore how the level of similarity of action choices between the automation and the human operator affects the resulting system's performance. A design process model for creating adaptive automation systems is complete, and was used to design an adaptive automation research environment. Data gathered using this system will be used to automate user task performance in the system, and allow for research into the effects of that automation.

This paper extends the concept of trust modeling within a multi-agent environment.  Trust modeling often focuses on identifying the appropriate trust level for the other agents in the environment and then using these levels to determine how to interact with each agent.  However, this type of modeling does not account for sneaky agents who are willing to cooperate when the stakes are low and take selfish, greedy actions when the rewards rise.  Adding trust to an interactive partially observable Markov decision process (I-POMDP) allows trust levels to be continuously monitored and corrected enabling agents to make better decisions.  The addition of trust modeling increases the decision process calculations, but solves more complex trust problems that are representative of the human world.  The modified I-POMDP reward function and belief models can be used to accurately track the trust levels of agents with hidden agendas.  Testing demonstrates that agents quickly identify the hidden trust levels to mitigate the impact of a deceitful agent.

Highly capable multiple robot architectures often resort to micromanagement to provide enhanced cooperative abilities, sacrificing individual autonomy. Conversely, multi-robot architectures that maintain individual autonomy are often limited in their cooperative abilities.  This article presents a modified three layer architecture that solves both of these issues.  The addition of a Coordinator layer to a three-layered approach provides a platform-independent interface for coordination on tasks and takes advantage of individual autonomy to improve coordination capabilities.  This reduces communication overhead versus many multi-robot architecture designs and allows for more straightforward resizing of the robot collective and increased individual autonomy.