Generative modeling of complex behaviors from labeled datasets has been a longstanding problem in decision making. Unlike language or image generation, decision making requires modeling actions - continuous-valued vectors that are multimodal in their distribution, potentially drawn from uncurated sources, where generation errors can compound in sequential prediction. A recent class of models called Behavior Transformers (BeT) addresses this by discretizing actions using k-means clustering to capture different modes. However, k-means struggles to scale for high-dimensional action spaces or long sequences, and lacks gradient information, and thus BeT suffers in modeling long-range actions. In this work, we present Vector-Quantized Behavior Transformer (VQ-BeT), a versatile model for behavior generation that handles multimodal action prediction, conditional generation, and partial observations. VQ-BeT augments BeT by tokenizing continuous actions with a hierarchical vector quantization module. Across seven environments including simulated manipulation, autonomous driving, and robotics, VQ-BeT improves on state-of-the-art models such as BeT and Diffusion Policies. Importantly, we demonstrate VQ-BeT's improved ability to capture behavior modes while accelerating inference speed 5x over Diffusion Policies. Videos and code can be found https://sjlee.cc/vq-bet

Achieving human-AI alignment in complex multi-agent games is crucial for creating trustworthy AI agents that enhance gameplay. We propose a method to evaluate this alignment using an interpretable task-sets framework, focusing on high-level behavioral tasks instead of low-level policies. Our approach has three components. First, we analyze extensive human gameplay data from Xbox's Bleeding Edge (100K+ games), uncovering behavioral patterns in a complex task space. This task space serves as a basis set for a behavior manifold capturing interpretable axes: fight-flight, explore-exploit, and solo-multi-agent. Second, we train an AI agent to play Bleeding Edge using a Generative Pretrained Causal Transformer and measure its behavior. Third, we project human and AI gameplay to the proposed behavior manifold to compare and contrast. This allows us to interpret differences in policy as higher-level behavioral concepts, e.g., we find that while human players exhibit variability in fight-flight and explore-exploit behavior, AI players tend towards uniformity. Furthermore, AI agents predominantly engage in solo play, while humans often engage in cooperative and competitive multi-agent patterns. These stark differences underscore the need for interpretable evaluation, design, and integration of AI in human-aligned applications. Our study advances the alignment discussion in AI and especially generative AI research, offering a measurable framework for interpretable human-agent alignment in multiplayer gaming.

Hierarchical reinforcement learning (HRL) incorporates temporal abstraction into reinforcement learning (RL) by explicitly taking advantage of hierarchical structure. Modern HRL typically designs a hierarchical agent composed of a high-level policy and low-level policies. The high-level policy selects which low-level policy to activate at a lower frequency and the activated low-level policy selects an action at each time step. Recent HRL algorithms have achieved performance gains over standard RL algorithms in synthetic navigation tasks. However, we cannot apply these HRL algorithms to real-world navigation tasks. One of the main challenges is that real-world navigation tasks require an agent to perform safe and interactive behaviors in dynamic environments. In this paper, we propose imagination-augmented HRL (IAHRL) that efficiently integrates imagination into HRL to enable an agent to learn safe and interactive behaviors in real-world navigation tasks. Imagination is to predict the consequences of actions without interactions with actual environments. The key idea behind IAHRL is that the low-level policies imagine safe and structured behaviors, and then the high-level policy infers interactions with surrounding objects by interpreting the imagined behaviors. We also introduce a new attention mechanism that allows our high-level policy to be permutation-invariant to the order of surrounding objects and to prioritize our agent over them. To evaluate IAHRL, we introduce five complex urban driving tasks, which are among the most challenging real-world navigation tasks. The experimental results indicate that IAHRL enables an agent to perform safe and interactive behaviors, achieving higher success rates and lower average episode steps than baselines.

Human behavior is inherently hierarchical, resulting from the decomposition of a task into subtasks or an abstract action into concrete actions. However, behavior is typically measured as a sequence of actions, which makes it difficult to infer its hierarchical structure. In this paper, we explore how people form hierarchically-structured plans, using an experimental paradigm that makes hierarchical representations observable: participants create programs that produce sequences of actions in a language with explicit hierarchical structure. This task lets us test two well-established principles of human behavior: utility maximization (i.e. using fewer actions) and minimum description length (MDL; i.e. having a shorter program). We find that humans are sensitive to both metrics, but that both accounts fail to predict a qualitative feature of human-created programs, namely that people prefer programs with reuse over and above the predictions of MDL. We formalize this preference for reuse by extending the MDL account into a generative model over programs, modeling hierarchy choice as the induction of a grammar over actions. Our account can explain the preference for reuse and provides the best prediction of human behavior, going beyond simple accounts of compressibility to highlight a principle that guides hierarchical planning.

Temporal abstraction and efficient planning pose significant challenges in offline reinforcement learning, mainly when dealing with domains that involve temporally extended tasks and delayed sparse rewards. Existing methods typically plan in the raw action space and can be inefficient and inflexible. Latent action spaces offer a more flexible paradigm, capturing only possible actions within the behavior policy support and decoupling the temporal structure between planning and modeling. However, current latent-action-based methods are limited to discrete spaces and require expensive planning. This paper presents a unified framework for continuous latent action space representation learning and planning by leveraging latent, score-based diffusion models. We establish the theoretical equivalence between planning in the latent action space and energy-guided sampling with a pretrained diffusion model and incorporate a novel sequence-level exact sampling method. Our proposed method, $\texttt{LatentDiffuser}$, demonstrates competitive performance on low-dimensional locomotion control tasks and surpasses existing methods in higher-dimensional tasks.

Reinforcement learning often needs to deal with the exponential growth of states and actions when exploring optimal control in high-dimensional spaces (often known as the curse of dimensionality). In this work, we address this issue by learning the inherent structure of action-wise similar MDP to appropriately balance the performance degradation versus sample/computational complexity. In particular, we partition the action spaces into multiple groups based on the similarity in transition distribution and reward function, and build a linear decomposition model to capture the difference between the intra-group transition kernel and the intra-group rewards. Both our theoretical analysis and experiments reveal a \emph{surprising and counter-intuitive result}: while a more refined grouping strategy can reduce the approximation error caused by treating actions in the same group as identical, it also leads to increased estimation error when the size of samples or the computation resources is limited. This finding highlights the grouping strategy as a new degree of freedom that can be optimized to minimize the overall performance loss. To address this issue, we formulate a general optimization problem for determining the optimal grouping strategy, which strikes a balance between performance loss and sample/computational complexity. We further propose a computationally efficient method for selecting a nearly-optimal grouping strategy, which maintains its computational complexity independent of the size of the action space.

Planning-based reinforcement learning has shown strong performance in tasks in discrete and low-dimensional continuous action spaces. However, planning usually brings significant computational overhead for decision-making, and scaling such methods to high-dimensional action spaces remains challenging. To advance efficient planning for high-dimensional continuous control, we propose Trajectory Autoencoding Planner (TAP), which learns low-dimensional latent action codes with a state-conditional VQ-VAE. The decoder of the VQ-VAE thus serves as a novel dynamics model that takes latent actions and current state as input and reconstructs long-horizon trajectories. During inference time, given a starting state, TAP searches over discrete latent actions to find trajectories that have both high probability under the training distribution and high predicted cumulative reward. Empirical evaluation in the offline RL setting demonstrates low decision latency which is indifferent to the growing raw action dimensionality. For Adroit robotic hand manipulation tasks with high-dimensional continuous action space, TAP surpasses existing model-based methods by a large margin and also beats strong model-free actor-critic baselines.

Reduced models are simplified versions of a given domain, designed to accelerate the planning process. Interest in reduced models has grown since the surprising success of determinization in the first international probabilistic planning competition, leading to the development of several enhanced determinization techniques. To address the drawbacks of previous determinization methods, we introduce a family of reduced models in which probabilistic outcomes are classified as one of two types: primary and exceptional. In each model that belongs to this family of reductions, primary outcomes can occur an unbounded number of times per trajectory, while exceptions can occur at most a finite number of times, specified by a parameter. Distinct reduced models are characterized by two parameters: the maximum number of primary outcomes per action, and the maximum number of occurrences of exceptions per trajectory. This family of reductions generalizes the well-known most-likely-outcome determinization approach, which includes one primary outcome per action and zero exceptional outcomes per plan. We present a framework to determine the benefits of planning with reduced models, and develop a continual planning approach that handles situations where the number of exceptions exceeds the specified bound during plan execution. Using this framework, we compare the performance of various reduced models and consider the challenge of generating good ones automatically. We show that each one of the dimensions---allowing more than one primary outcome or planning for some limited number of exceptions---could improve performance relative to standard determinization. The results place previous work on determinization in a broader context and lay the foundation for a systematic exploration of the space of model reductions.

Explainability in Artificial Intelligence has been revived as a topic of active research by the need of conveying safety and trust to users in the `how' and `why' of automated decision-making. Whilst a plethora of approaches have been developed for post-hoc explainability, only a few focus on how to use domain knowledge, and how this influences the understandability of an explanation from the users' perspective. In this paper we show how ontologies help the understandability of interpretable machine learning models, such as decision trees. In particular, we build on Trepan, an algorithm that explains artificial neural networks by means of decision trees, and we extend it to include ontologies modeling domain knowledge in the process of generating explanations. We present the results of a user study that measures the understandability of decision trees in domains where explanations are critical, namely, in finance and medicine. Our study shows that decision trees taking into account domain knowledge during generation are more understandable than those generated without the use of ontologies.

Compact representations of objects is a common concept in computer science. Automated planning can be viewed as a case of this concept: a planning instance is a compact implicit representation of a graph and the problem is to find a path (a plan) in this graph. While the graphs themselves are represented compactly as planning instances, the paths are usually represented explicitly as sequences of actions. Some cases are known where the plans always have compact representations, for example, using macros. We show that these results do not extend to the general case, by proving a number of bounds for compact representations of plans under various criteria, like efficient sequential or random access of actions. In addition to this, we show that our results have consequences for what can be gained from reformulating planning into some other problem. As a contrast to this we also prove a number of positive results, demonstrating restricted cases where plans do have useful compact representations, as well as proving that macro plans have favourable access properties. Our results are finally discussed in relation to other relevant contexts.