Jonathan Balloch, Ph.D.

Reinforcement learning (RL) using world models has found significant recent successes. However, when a sudden change to world mechanics or properties occurs then agent performance and reliability can dramatically decline. We refer to the sudden change in visual properties or state transitions as novelties. Implementing novelty detection within generated world model frameworks is a crucial task for protecting the agent when deployed. In this paper, we propose straightforward bounding approaches to incorporate novelty detection into world model RL agents, by utilizing the misalignment of the world model's hallucinated states and the true observed states as an anomaly score. We provide effective approaches to detecting novelties in a distribution of transitions learned by an agent in a world model. Finally, we show the advantage of our work in a novel environment compared to traditional machine learning novelty detection methods as well as currently accepted RL focused novelty detection algorithms.

Real-world autonomous decision-making systems, from robots to recommendation engines, must operate in environments that change over time. While deep reinforcement learning (RL) has shown an impressive ability to learn optimal policies in stationary environments, most methods are data intensive and assume a world that does not change between training and test time. As a result, conventional RL methods struggle to adapt when conditions change. This poses a fundamental challenge: how can RL agents efficiently adapt their behavior when encountering novel environmental changes during deployment without catastrophically forgetting useful prior knowledge? This dissertation demonstrates that efficient online adaptation requires two key capabilities: (1) prioritized exploration and sampling strategies that help identify and learn from relevant experiences, and (2) selective preservation of prior knowledge through structured representations that can be updated without disruption to reusable components.

In deep reinforcement learning (RL) research, there has been a concerted effort to design more efficient and productive exploration methods while solving sparse-reward problems. These exploration methods often share common principles (e.g., improving diversity) and implementation details (e.g., intrinsic reward). Prior work found that non-stationary Markov decision processes (MDPs) require exploration to efficiently adapt to changes in the environment with online transfer learning. However, the relationship between specific exploration characteristics and effective transfer learning in deep RL has not been characterized. In this work, we seek to understand the relationships between salient exploration characteristics and improved performance and efficiency in transfer learning. We test eleven popular exploration algorithms on a variety of transfer types -- or "novelties" -- to identify the characteristics that positively affect online transfer learning. Our analysis shows that some characteristics correlate with improved performance and efficiency across a wide range of transfer tasks, while others only improve transfer performance with respect to specific environment changes. From our analysis, we make recommendations about which exploration algorithm characteristics are best suited to specific transfer situations.

Open-world novelty--a sudden change in the mechanics or properties of an environment--is a common occurrence in the real world. Novelty adaptation is an agent's ability to improve its policy performance post-novelty. Most reinforcement learning (RL) methods assume that the world is a closed, fixed process. Consequentially, RL policies adapt inefficiently to novelties. To address this, we introduce WorldCloner, an end-to-end trainable neuro-symbolic world model for rapid novelty adaptation. WorldCloner learns an efficient symbolic representation of the pre-novelty environment transitions, and uses this transition model to detect novelty and efficiently adapt to novelty in a single-shot fashion. Additionally, WorldCloner augments the policy learning process using imagination-based adaptation, where the world model simulates transitions of the post-novelty environment to help the policy adapt. By blending "imagined" transitions with interactions in the post-novelty environment, performance can be recovered with fewer total environment interactions. Using environments designed for studying novelty in sequential decision-making problems, we show that the symbolic world model helps its neural policy adapt more efficiently than model-based and model-based neural-only reinforcement learning methods.

A robust body of reinforcement learning techniques have been developed to solve complex sequential decision making problems. However, these methods assume that train and evaluation tasks come from similarly or identically distributed environments. This assumption does not hold in real life where small novel changes to the environment can make a previously learned policy fail or introduce simpler solutions that might never be found. To that end we explore the concept of {\em novelty}, defined in this work as the sudden change to the mechanics or properties of environment. We provide an ontology of for novelties most relevant to sequential decision making, which distinguishes between novelties that affect objects versus actions, unary properties versus non-unary relations, and the distribution of solutions to a task. We introduce NovGrid, a novelty generation framework built on MiniGrid, acting as a toolkit for rapidly developing and evaluating novelty-adaptation-enabled reinforcement learning techniques. Along with the core NovGrid we provide exemplar novelties aligned with our ontology and instantiate them as novelty templates that can be applied to many MiniGrid-compliant environments. Finally, we present a set of metrics built into our framework for the evaluation of novelty-adaptation-enabled machine-learning techniques, and show characteristics of a baseline RL model using these metrics.

Rehearsal is a critical component for class-incremental continual learning, yet it requires a substantial memory budget. Our work investigates whether we can significantly reduce this memory budget by leveraging unlabeled data from an agent's environment in a realistic and challenging continual learning paradigm. Specifically, we explore and formalize a novel semi-supervised continual learning (SSCL) setting, where labeled data is scarce yet non-i.i.d. unlabeled data from the agent's environment is plentiful. Importantly, data distributions in the SSCL setting are realistic and therefore reflect object class correlations between, and among, the labeled and unlabeled data distributions. We show that a strategy built on pseudo-labeling, consistency regularization, Out-of-Distribution (OoD) detection, and knowledge distillation reduces forgetting in this setting. Our approach, DistillMatch, increases performance over the state-of-the-art by no less than 8.7% average task accuracy and up to 54.5% average task accuracy in SSCL CIFAR-100 experiments. Moreover, we demonstrate that DistillMatch can save up to 0.23 stored images per processed unlabeled image compared to the next best method which only saves 0.08. Our results suggest that focusing on realistic correlated distributions is a significantly new perspective, which accentuates the importance of leveraging the world's structure as a continual learning strategy.

MacGyvering is defined as creating or repairing something in an inventive or improvised way by utilizing objects that are available at hand. In this paper, we explore a subset of Macgyvering problems involving tool construction, i.e., creating tools from parts available in the environment. We formalize the overall problem domain of tool Macgyvering, introducing three levels of complexity for tool construction and substitution problems, and presenting a novel computational framework aimed at solving one level of the tool Macgyvering problem, specifically contributing a novel algorithm for tool construction based on geometric reasoning. We validate our approach by constructing three tools using a 7-DOF robot arm.

The ability of a robot to reason about the geometry and semantics of its environment is fundamental to interactive robot behaviors, but often challenging due to perception frameworks that are trained on too little data or data not representative of the robots environment. In this work, we investigate the potential gains of using synthetic data to augment the training process of convolutional neural networks designed to enable real-time semantic segmentation for robots with limited real-world training data. We investigate the degree to which a larger amounts of data improves performance when training such a model, the relationship between the way a deep neural network is trained using multiple sources of synthetic segmentation data to pretrain standard segmentation datasets that apply to robotics and autonomous driving, and show that our method outperforms both training from scratch and standard data augmentation practices like pretraining on ImageNet. We show that synthetic data does continue to improve these models in spite of real-time model architectures having many fewer parameters than typical deep neural networks, and therefore hypothetically less representational power. Finally, we show how this approach generalizes to small purpose-built robot vision datasets on data acquired using an HRI robot.