Abstract There is a gap between current methods that explain deep learning models that work on imbalanced image data and the needs of the imbalanced learning community. Existing methods that explain imbalanced data are geared toward binary classification, single layer machine learning models and low dimensional data. Current eXplainable Artificial Intelligence (XAI) techniques for vision data mainly focus on mapping predictions of specific instances to inputs, instead of examining global data properties and complexities of entire classes. Therefore, there is a need for a framework that is tailored to modern deep networks, that incorporates large, high dimensional, multi-class datasets, and uncovers data complexities commonly found in imbalanced data. We propose a set of techniques that can be used by both deep learning model users to identify, visualize and understand class prototypes, sub-concepts and outlier instances; and by imbalanced learning algorithm developers to detect features and class exemplars that are key to model performance. The components of our framework can be applied sequentially in their entirety or individually, making it fully flexible to the user’s specific needs ( https://github.com/dd1github/XAI_for_Imbalanced_Learning ).

Common approaches to learn complex tasks in reinforcement learning include reward shaping, environmental hints, or a curriculum. Yet few studies examine how they compare to each other, when one might prefer one approach, or how they may complement each other. As a first step in this direction, we compare reward shaping, hints, and curricula for a Deep RL agent in the game of Minecraft. We seek to answer whether reward shaping, visual hints, or the curricula have the most impact on performance, which we measure as the time to reach the target, the distance from the target, the cumulative reward, or the number of actions taken. Our analyses show that performance is most impacted by the curriculum used and visual hints; shaping had less impact. For similar navigation tasks, the results suggest that designing an effective curriculum and providing appropriate hints most improve the performance. Common approaches to learn complex tasks in reinforcement learning include reward shaping, environmental hints, or a curriculum, yet few studies examine how they compare to each other. We compare these approaches for a Deep RL agent in the game of Minecraft and show performance is most impacted by the curriculum used and visual hints; shaping had less impact. For similar navigation tasks, this suggests that designing an effective curriculum with hints most improve the performance.

Sci-fi narratives permeating the collective consciousness endow AI Rebellion with ample negative connotations. However, for AI agents, as for humans, attitudes of protest, objection, and rejection have many potential benefits in support of ethics, safety, self-actualization, solidarity, and social justice, and are necessary in a wide variety of contexts. We launch a conversation on constructive AI rebellion and describe a framework meant to support discussion, implementation, and deployment of AI Rebel Agents as protagonists of positive narratives.

Heuristics serve as a powerful tool in modern domain-independent planning (DIP) systems by providing critical guidance during the search for high-quality solutions. However, they have not been broadly used with hierarchical planning techniques, which are more expressive and tend to scale better in complex domains by exploiting additional domain-specific knowledge. Complicating matters, we show that for Hierarchical Goal Network (HGN) planning, a goal-based hierarchical planning formalism that we focus on in this paper, any poly-time heuristic that is derived from a delete-relaxation DIP heuristic has to make some relaxation of the hierarchical semantics. To address this, we present a principled framework for incorporating DIP heuristics into HGN planning using a simple relaxation of the HGN semantics we call Hierarchy-Relaxation. This framework allows for computing heuristic estimates of HGN problems using any DIP heuristic in an admissibility-preserving manner. We demonstrate the feasibility of this approach by using the LMCut heuristic to guide an optimal HGN planner. Our empirical results with three benchmark domains demonstrate that simultaneously leveraging hierarchical knowledge and heuristic guidance substantially improves planning performance.

We describe the Tactical Battle Manager (TBM), an intelligent agent that uses several integrated artificial intelligence techniques to control an autonomous unmanned aerial vehicle in simulated beyond-visual-range (BVR) air combat scenarios. The TBM incorporates goal reasoning, automated planning, opponent behavior recognition, state prediction, and discrepancy detection to operate in a real-time, dynamic, uncertain, and adversarial environment. We describe evidence from our empirical study that the TBM significantly outperforms an expert-scripted agent in BVR scenarios. We also report the results of an ablation study which indicates that all components of our agent architecture are needed to maximize mission performance.

There is an impressive body of work on developing heuristics and other reasoning algorithms to guide search in optimal and anytime planning algorithms for classical planning. However, very little effort has been directed towards developing analogous techniques to guide search towards high-quality solutions in hierarchical planning formalisms like HTN planning, which allows using additional domain-specific procedural control knowledge. In lieu of such techniques, this control knowledge often needs to provide the necessary search guidance to the planning algorithm, which imposes a substantial burden on the domain author and can yield brittle or error-prone domain models. We address this gap by extending recent work on a new hierarchical goal-based planning formalism called Hierarchical Goal Network (HGN) Planning to develop the Hierarchically-Optimal Goal Decomposition Planner (HOpGDP), an HGN planning algorithm that computes hierarchically-optimal plans. HOpGDP is guided by $h_{HL}$, a new HGN planning heuristic that extends existing admissible landmark-based heuristics from classical planning to compute admissible cost estimates for HGN planning problems. Our experimental evaluation across three benchmark planning domains shows that HOpGDP compares favorably to both optimal classical planners due to its ability to use domain-specific procedural knowledge, and a blind-search version of HOpGDP due to the search guidance provided by $h_{HL}$.

Many analysis tasks involve linked nodes, such as people connected by friendship links. Research on link-based classification (LBC) has studied how to leverage these connections to improve classification accuracy. Most such prior research has assumed the provision of a densely labeled training network. Instead, this article studies the common and challenging case when LBC must use a single sparsely labeled network for both learning and inference, a case where existing methods often yield poor accuracy. To address this challenge, we introduce a novel method that enables prediction via “neighbor attributes,” which were briefly considered by early LBC work but then abandoned due to perceived problems. We then explain, using both extensive experiments and loss decomposition analysis, how using neighbor attributes often significantly improves accuracy. We further show that using appropriate semi-supervised learning (SSL) is essential to obtaining the best accuracy in this domain and that the gains of neighbor attributes remain across a range of SSL choices and data conditions. Finally, given the challenges of label sparsity for LBC and the impact of neighbor attributes, we show that multiple previous studies must be re-considered, including studies regarding the best model features, the impact of noisy attributes, and strategies for active learning.

Hierarchical Task Network (HTN) planning with task insertion (TIHTN planning) is a variant of HTN planning. In HTN planning, the only means to alter task networks is to decompose compound tasks. In TIHTN planning, tasks may also be inserted directly. In this paper we provide tight complexity bounds for TIHTN planning along two axis: whether variables are allowed and whether methods must be totally ordered.

Agents with incomplete environment models are likely to be surprised, and this represents an opportunity to learn. We investigate approaches for situated agents to detect surprises, discriminate among different forms of surprise, and hypothesize new models for the unknown events that surprised them. We instantiate these approaches in a new goal reasoning agent (named FoolMeTwice), investigate its performance in simulation studies, and report that it produces plans with significantly reduced execution cost in comparison to not learning models for surprising events.

Many classification tasks involve linked nodes, such as people connected by friendship links. For such networks, accuracy might be increased by including, for each node, the (a) labels or (b) attributes of neighboring nodes as model features. Recent work has focused on option (a), because early work showed it was more accurate and because option (b) fit poorly with discriminative classifiers. We show, however, that when the network is sparsely labeled, "relational classification" based on neighbor attributes often has higher accuracy than "collective classification" based on neighbor labels. Moreover, we introduce an efficient method that enables discriminative classifiers to be used with neighbor attributes, yielding further accuracy gains. We show that these effects are consistent across a range of datasets, learning choices, and inference algorithms, and that using both neighbor attributes and labels often produces the best accuracy.

To operate autonomously in complex environments, an agent must monitor its environment and determine how to respond to new situations. To be considered intelligent, an agent should select actions in pursuit of its goals, and adapt accordingly when its goals need revision. However, most agents assume that their goals are given to them; they cannot recognize when their goals should change. Thus, they have difficulty coping with the complex environments of strategy simulations that are continuous, partially observable, dynamic, and open with respect to new objects. To increase intelligent agent autonomy, we are investigating a conceptual model for goal reasoning called Goal‐Driven Autonomy (GDA), which allows agents to generate and reason about their goals in response to environment changes. Our hypothesis is that GDA enables an agent to respond more effectively to unexpected events in complex environments. We instantiate the GDA model in ARTUE (Autonomous Response to Unexpected Events), a domain‐independent autonomous agent. We evaluate ARTUE on scenarios from two complex strategy simulations, and report on its comparative benefits and limitations. By employing goal reasoning, ARTUE outperforms an off‐line planner and a discrepancy‐based replanner on scenarios requiring reasoning about unobserved objects and facts and on scenarios presenting opportunities outside the scope of its current mission.

Relational data representations have become an increasingly important topic due to the recent proliferation of network datasets (e.g., social, biological, information networks) and a corresponding increase in the application of Statistical Relational Learning (SRL) algorithms to these domains. In this article, we examine and categorize techniques for transforming graph-based relational data to improve SRL algorithms. In particular, appropriate transformations of the nodes, links, and/or features of the data can dramatically affect the capabilities and results of SRL algorithms. We introduce an intuitive taxonomy for data representation transformations in relational domains that incorporates link transformation and node transformation as symmetric representation tasks. More specifically, the transformation tasks for both nodes and links include (i) predicting their existence, (ii) predicting their label or type, (iii) estimating their weight or importance, and (iv) system- atically constructing their relevant features. We motivate our taxonomy through detailed examples and use it to survey competing approaches for each of these tasks. We also dis- cuss general conditions for transforming links, nodes, and features. Finally, we highlight challenges that remain to be addressed.

We organized a data mining challenge in “unsupervised and transfer learning” (the UTL challenge), in collaboration with the DARPA Deep Learning program. The goal of this year's challenge was to learn good data representations that can be re-used across tasks by building models that capture regularities of the input space. The representations provided by the participants were evaluated by the organizers on supervised learning “target tasks”, which were unknown to the participants. In a first phase of the challenge, the competitors were given only unlabeled data to learn their data representation. In a second phase of the challenge, the competitors were also provided with a limited amount of labeled data from “source tasks”, distinct from the “target tasks”. We made available large datasets from various application domains: handwriting recognition, image recognition, video processing, text processing, and ecology. The results indicate that learned data representation yield results significantly better than what can be achieved with raw data or data preprocessed with standard normalizations and functional transforms. The UTL challenge is part of the IJCNN 2011 competition program1. The website of the challenge remains open for submission of new methods beyond the termination of the challenge as a resource for students and researchers2.

Planning in dynamic continuous environments requires reasoning about nonlinear continuous effects, which previous Hierarchical Task Network (HTN) planners do not support. In this paper, we extend an existing HTN planner with a new state projection algorithm. To our knowledge, this is the first HTN planner that can reason about nonlinear continuous effects. We use a wait action to instruct this planner to consider continuous effects in a given state. We also introduce a new planning domain to demonstrate the benefits of planning with nonlinear continuous effects. We compare our approach with a linear continuous effects planner and a discrete effects HTN planner on a benchmark domain, which reveals that its additional costs are largely mitigated by domain knowledge. Finally, we present an initial application of this algorithm in a practical domain, a Navy training simulation, illustrating the utility of this approach for planning in dynamic continuous environments.

Modern complex games and simulations pose many challenges for an intelligent agent, including partial observability, continuous time and effects, hostile opponents, and exogenous events. We present ARTUE (Autonomous Response to Unexpected Events), a domain-independent autonomous agent that dynamically reasons about what goals to pursue in response to unexpected circumstances in these types of environments. ARTUE integrates AI research in planning, environment monitoring, explanation, goal generation, and goal management. To explain our conceptualization of the problem ARTUE addresses, we present a new conceptual framework, goal-driven autonomy, for agents that reason about their goals. We evaluate ARTUE on scenarios in the TAO Sandbox, a Navy training simulation, and demonstrate its novel architecture, which includes components for Hierarchical Task Network planning, explanation, and goal management. Our evaluation shows that ARTUE can perform well in a complex environment and that each component is necessary and contributes to the performance of the integrated system.

A significant challenge in developing planning systems for practical applications is the difficulty of acquiring the domain knowledge needed by such systems. One method for acquiring this knowledge is to learn it from plan traces, but this method typically requires a huge number of plan traces to converge. In this paper, we show that the problem with slow convergence can be circumvented by having the learner generate solution plans even before the planning domain is completely learned. Our empirical results show that these improvements reduce the size of the training set that is needed to find correct answers to a large percentage of planning problems in the test set. 1.

A great challenge in developing planning systems for practical applications is the difficulty of acquiring the domain information needed to guide such systems. This paper describes a way to learn some of that knowledge. More specifically, the following points are discussed. (1) We introduce a theoretical basis for formally defining algorithms that learn preconditions for Hierarchical Task Network (HTN) methods. (2) We describe Candidate Elimination Method Learner (CaMeL), a supervised, eager, and incremental learning process for preconditions of HTN methods. We state and prove theorems about CaMeL's soundness, completeness, and convergence properties. (3) We present empirical results about CaMeL's convergence under various conditions. Among other things, CaMeL converges the fastest on the preconditions of the HTN methods that are needed the most often. Thus CaMeL's output can be useful even before it has fully converged.

We introduce LET (Lesson Elicitation Tool), which uses domain and linguistic knowledge to guide users during their submission of lessons learned. LET can detect a user's need for instructions and disambiguates expressions while collecting taxonomic domain knowledge.

Instance‐based representations have been applied to numerous classification tasks with some success. Most of these applications involved predicting a symbolic class based on observed attributes. This paper presents an instance‐based method for predicting a numeric value based on observed attributes. We prove that, given enough instances, if the numeric values are generated by continuous functions with bounded slope, then the predicted values are accurate approximations of the actual values. We demonstrate the utility of this approach by comparing it with a standard approach for value prediction. The instance‐based approach requires neither ad hoc parameters nor background knowledge.