AAAI 2023 Papers

Neural scene representation and rendering methods have shown promise in learning the implicit form of scene structure without supervision. However, the implicit representation learned in most existing methods is non-expandable and cannot be inferred online for novel scenes, which makes the learned representation difficult to be applied across different reinforcement learning (RL) tasks. In this work, we introduce Scene Memory Network (SMN) to achieve online spatial memory construction and expansion for view rendering in novel scenes. SMN models the camera projection and back-projection as spatially aware memory control processes, where the memory values store the information of the partial 3D area, and the memory keys indicate the position of that area. The memory controller can learn the geometry property from observations without the camera's intrinsic parameters and depth supervision. We further apply the memory constructed by SMN to exploration and navigation tasks. The experimental results reveal the generalization ability of our proposed SMN in large-scale scene synthesis and its potential to improve the performance of spatial RL tasks.

There has been a surge of interest in learning optimal decision trees using mixed-integer programs (MIP) in recent years, as heuristic-based methods do not guarantee optimality and find it challenging to incorporate constraints that are critical for many practical applications. However, existing MIP methods that build on an arc-based formulation do not scale well as the number of binary variables is in the order of 2 to the power of the depth of the tree and the size of the dataset. Moreover, they can only handle sample-level constraints and linear metrics. In this paper, we propose a novel path-based MIP formulation where the number of decision variables is independent of dataset size. We present a scalable column generation framework to solve the MIP. Our framework produces a multiway-split tree which is more interpretable than the typical binary-split trees due to its shorter rules. Our framework is more general as it can handle nonlinear metrics such as F1 score, and incorporate a broader class of constraints. We demonstrate its efficacy with extensive experiments. We present results on datasets containing up to 1,008,372 samples while existing MIP-based decision tree models do not scale well on data beyond a few thousand points. We report superior or competitive results compared to the state-of-art MIP-based methods with up to a 24X reduction in runtime.

Previous research on the comprehensive negotiation strategy using deep reinforcement learning (RL) has scalability issues of not performing effectively in the large-sized domains. We improve negotiation strategy via deep RL by considering an issue-based represented deep policy network to deal with multi-issue negotiation. The architecture of the proposed learning agent considers the characteristics of multi-issue negotiation domains and policy-based learning. We demonstrate that proposed method achieve equivalent or higher utility than existing negotiation agents in the large-sized domains.

Active Directory (AD) is the default security management system for Windows domain networks. An AD environment naturally describes an attack graph where nodes represent computers/accounts/security groups, and edges represent existing accesses/known exploits that allow the attacker to gain access from one node to another. Motivated by practical AD use cases, we study a Stackelberg game between one attacker and one defender. There are multiple entry nodes for the attacker to choose from and there is a single target (Domain Admin). Every edge has a failure rate. The attacker chooses the attack path with the maximum success rate. The defender can block a limited number of edges (i.e., revoke accesses) from a set of blockable edges, limited by budget. The defender's aim is to minimize the attacker's success rate. We exploit the tree-likeness of practical AD graphs to design scalable algorithms. We propose two novel methods that combine theoretical fixed parameter analysis and practical optimisation techniques. For graphs with small tree widths, we propose a tree decomposition based dynamic program. We then propose a general method for converting tree decomposition based dynamic programs to reinforcement learning environments, which leads to an anytime algorithm that scales better, but loses the optimality guarantee. For graphs with small numbers of non-splitting paths (a parameter we invent specifically for AD graphs), we propose a kernelization technique that significantly downsizes the model, which is then solved via mixed-integer programming. Experimentally, our algorithms scale to handle synthetic AD graphs with tens of thousands of nodes.

This paper studies restless multi-armed bandit (RMAB) problems with unknown arm transition dynamics but with known correlated arm features. The goal is to learn a model to predict transition dynamics given features, where the Whittle index policy solves the RMAB problems using predicted transitions. However, prior works often learn the model by maximizing the predictive accuracy instead of final RMAB solution quality, causing a mismatch between training and evaluation objectives. To address this shortcoming, we propose a novel approach for decision-focused learning in RMAB that directly trains the predictive model to maximize the Whittle index solution quality. We present three key contributions: (i) we establish differentiability of the Whittle index policy to support decision-focused learning; (ii) we significantly improve the scalability of decision-focused learning approaches in sequential problems, specifically RMAB problems; (iii) we apply our algorithm to a previously collected dataset of maternal and child health to demonstrate its performance. Indeed, our algorithm is the first for decision-focused learning in RMAB that scales to real-world problem sizes.

Meta-learning owns unique effectiveness and swiftness in tackling emerging tasks with limited data. Its broad applicability is revealed by viewing it as a bi-level optimization problem. The resultant algorithmic viewpoint however, faces scalability issues when the inner-level optimization relies on gradient-based iterations. Implicit differentiation has been considered to alleviate this challenge, but it is restricted to an isotropic Gaussian prior, and only favors deterministic meta-learning approaches. This work markedly mitigates the scalability bottleneck by cross-fertilizing the benefits of implicit differentiation to probabilistic Bayesian meta-learning. The novel implicit Bayesian meta-learning (iBaML) method not only broadens the scope of learnable priors, but also quantifies the associated uncertainty. Furthermore, the ultimate complexity is well controlled regardless of the inner-level optimization trajectory. Analytical error bounds are established to demonstrate the precision and efficiency of the generalized implicit gradient over the explicit one. Extensive numerical tests are also carried out to empirically validate the performance of the proposed method.

Over recent years, graph convolutional networks emerged as powerful node clustering methods and have set state of the art results for this task. In this paper, we argue that some of these methods are unnecessarily complex and propose a node clustering model that is more scalable while being more effective. The proposed model uses Laplacian smoothing to learn an initial representation of the graph before applying an efficient self-expressive subspace clustering procedure. This is performed via learning a factored coefficient matrix. These factors are then embedded into a new feature space in such a way as to generate a valid affinity matrix (symmetric and non-negative) on which an implicit spectral clustering algorithm is performed. Experiments on several real-world attributed datasets demonstrate the cost-effective nature of our method with respect to the state of the art.

It is now well known that neural networks can be wrong with high confidence in their predictions, leading to poor calibration. The most common post-hoc approach to compensate for this is to perform temperature scaling, which adjusts the confidences of the predictions on any input by scaling the logits by a fixed value. Whilst this approach typically improves the average calibration across the whole test dataset, this improvement typically reduces the individual confidences of the predictions irrespective of whether the classification of a given input is correct or incorrect. With this insight, we base our method on the observation that different samples contribute to the calibration error by varying amounts, with some needing to increase their confidence and others needing to decrease it. Therefore, for each input, we propose to predict a different temperature value, allowing us to adjust the mismatch between confidence and accuracy at a finer granularity. Our method is applied post-hoc, enabling it to be very fast with a negligible memory footprint and is applied to off-the-shelf pre-trained classifiers. We test our method on the ResNet50 and WideResNet28-10 architectures using the CIFAR10/100 and Tiny-ImageNet datasets, showing that producing per-data-point temperatures improves the expected calibration error across the whole test set.

Bayesian neural networks (BNNs) retain NN structures with a probability distribution placed over their weights. With the introduced uncertainties and redundancies, BNNs are proper choices of robust controllers for safety-critical control systems. This paper considers the problem of verifying the safety of nonlinear closed-loop systems with BNN controllers over unbounded-time horizon. In essence, we compute a safe weight set such that as long as the BNN controller is always applied with weights sampled from the safe weight set, the controlled system is guaranteed to be safe. We propose a novel two-phase method for the safe weight set computation. First, we construct a reference safe control set that constraints the control inputs, through polynomial approximation to the BNN controller followed by polynomial-optimization-based barrier certificate generation. Then, the computation of safe weight set is reduced to a range inclusion problem of the BNN on the system domain w.r.t. the safe control set, which can be solved incrementally and the set of safe weights can be extracted. Compared with the existing method based on invariant learning and mixed-integer linear programming, we could compute safe weight sets with larger radii on a series of linear benchmarks. Moreover, experiments on a series of widely used nonlinear control tasks show that our method can synthesize large safe weight sets with probability measure as high as 95% even for a large-scale system of dimension 7.

In recent years, learning-based autonomous systems have emerged as a promising tool for automating many crucial tasks. The key question is how we can build trust in such systems for safety-critical applications. My research aims to focus on the creation and validation of safety frameworks that leverage multiple sources of information. The ultimate goal is to establish a solid foundation for a long-term research program aimed at understanding the role of fidelity in simulators for safety validation and robot learning.

Neural network pruning is a technique of network compression by removing weights of lower importance from an optimized neural network. Often, pruned networks are compared in terms of accuracy, which is realized in terms of rewards for Deep Reinforcement Learning (DRL) networks. However, networks that estimate control actions for safety-critical tasks, must also adhere to safety requirements along with obtaining rewards. We propose a methodology to iteratively refine the weights of a pruned neural network such that we get a sparse high-performance network without significant side effects on safety.

Applications abound in which optimization problems must be repeatedly solved, each time with new (but similar) data. Analytic optimization algorithms can be hand-designed to provably solve these problems in an iterative fashion. On one hand, data-driven algorithms can "learn to optimize" (L2O) with much fewer iterations and similar cost per iteration as general-purpose optimization algorithms. On the other hand, unfortunately, many L2O algorithms lack converge guarantees. To fuse the advantages of these approaches, we present a Safe-L2O framework. Safe-L2O updates incorporate a safeguard to guarantee convergence for convex problems with proximal and/or gradient oracles. The safeguard is simple and computationally cheap to implement, and it is activated only when the data-driven L2O updates would perform poorly or appear to diverge. This yields the numerical benefits of employing machine learning to create rapid L2O algorithms while still guaranteeing convergence. Our numerical examples show convergence of Safe-L2O algorithms, even when the provided data is not from the distribution of training data.

Traffic signal control is safety-critical for our daily life. Roughly one-quarter of road accidents in the U.S. happen at intersections due to problematic signal timing, urging the development of safety-oriented intersection control. However, existing studies on adaptive traffic signal control using reinforcement learning technologies have focused mainly on minimizing traffic delay but neglecting the potential exposure to unsafe conditions. We, for the first time, incorporate road safety standards as enforcement to ensure the safety of existing reinforcement learning methods, aiming toward operating intersections with zero collisions. We have proposed a safety-enhanced residual reinforcement learning method (SafeLight) and employed multiple optimization techniques, such as multi-objective loss function and reward shaping for better knowledge integration. Extensive experiments are conducted using both synthetic and real-world benchmark datasets. Results show that our method can significantly reduce collisions while increasing traffic mobility.

Safe exploration is a common problem in reinforcement learning (RL) that aims to prevent agents from making disastrous decisions while exploring their environment. A family of approaches to this problem assume domain knowledge in the form of a (partial) model of this environment to decide upon the safety of an action. A so-called shield forces the RL agent to select only safe actions. However, for adoption in various applications, one must look beyond enforcing safety and also ensure the applicability of RL with good performance. We extend the applicability of shields via tight integration with state-of-the-art deep RL, and provide an extensive, empirical study in challenging, sparse-reward environments under partial observability. We show that a carefully integrated shield ensures safety and can improve the convergence rate and final performance of RL agents. We furthermore show that a shield can be used to bootstrap state-of-the-art RL agents: they remain safe after initial learning in a shielded setting, allowing us to disable a potentially too conservative shield eventually.

We study safe policy improvement (SPI) for partially observable Markov decision processes (POMDPs). SPI is an offline reinforcement learning (RL) problem that assumes access to (1) historical data about an environment, and (2) the so-called behavior policy that previously generated this data by interacting with the environment. SPI methods neither require access to a model nor the environment itself, and aim to reliably improve upon the behavior policy in an offline manner. Existing methods make the strong assumption that the environment is fully observable. In our novel approach to the SPI problem for POMDPs, we assume that a finite-state controller (FSC) represents the behavior policy and that finite memory is sufficient to derive optimal policies. This assumption allows us to map the POMDP to a finite-state fully observable MDP, the history MDP. We estimate this MDP by combining the historical data and the memory of the FSC, and compute an improved policy using an off-the-shelf SPI algorithm. The underlying SPI method constrains the policy space according to the available data, such that the newly computed policy only differs from the behavior policy when sufficient data is available. We show that this new policy, converted into a new FSC for the (unknown) POMDP, outperforms the behavior policy with high probability. Experimental results on several well-established benchmarks show the applicability of the approach, even in cases where finite memory is not sufficient.

Multi-view deep classification expects to obtain better classification performance than using a single view. However, due to the uncertainty and inconsistency of data sources, adding data views does not necessarily lead to the performance improvements in multi-view classification. How to avoid worsening classification performance when adding views is crucial for multi-view deep learning but rarely studied. To tackle this limitation, in this paper, we reformulate the multi-view classification problem from the perspective of safe learning and thereby propose a Safe Multi-view Deep Classification (SMDC) method, which can guarantee that the classification performance does not deteriorate when fusing multiple views. In the SMDC method, we dynamically integrate multiple views and estimate the inherent uncertainties among multiple views with different root causes based on evidence theory. Through minimizing the uncertainties, SMDC promotes the evidences from data views for correct classification, and in the meantime excludes the incorrect evidences to produce the safe multi-view classification results. Furthermore, we theoretically prove that in the safe multi-view classification, adding data views will certainly not increase the empirical risk of classification. The experiments on various kinds of multi-view datasets validate that the proposed SMDC method can achieve precise and safe classification results.

Safe Interval Path Planning (SIPP) is a powerful algorithm for solving a single-agent pathfinding problem where the agent is confined to a graph and certain vertices/edges of this graph are blocked at certain time intervals due to dynamic obstacles that populate the environment. The original SIPP algorithm relies on the assumption that the agent is able to stop instantaneously. However, this assumption often does not hold in practice, e.g. a mobile robot moving at a cruising speed cannot stop immediately but rather requires gradual deceleration to a full stop that takes time. In other words, the robot is subject to kinodynamic constraints. Unfortunately, as we show in this work, in such a case, the original SIPP is incomplete. To this end, we introduce a novel variant of SIPP that is provably complete and optimal for planning with acceleration/deceleration. In the experimental evaluation, we show that the key property of the original SIPP still holds for the modified version: it performs much fewer expansions compared to A* and, as a result, is notably faster.

It is envisioned that in the near future autonomous systems such as multi-agent systems, will co-exist with humans, e.g., autonomous vehicles will share roads with human drivers. These safety-critical scenarios require formally provable safety guarantees so that the robots will never collide with humans or with each other. It is challenging to provide such guarantees in the real world due to the stochastic environments and inaccurate models of heterogeneous agents including robots and humans. My PhD research investigates decision-making algorithm design for provably-correct safety guarantees in mixed multi-agent systems.

In practice, Wearable Human Activity Recognition (WHAR) models usually face performance degradation on the new user due to user variance. Unsupervised domain adaptation (UDA) becomes the natural solution to cross-user WHAR under annotation scarcity. Existing UDA models usually align samples across domains without differentiation, which ignores the difference among samples. In this paper, we propose an unsupervised domain adaptation model with sample weight learning (SWL-Adapt) for cross-user WHAR. SWL-Adapt calculates sample weights according to the classification loss and domain discrimination loss of each sample with a parameterized network. We introduce the meta-optimization based update rule to learn this network end-to-end, which is guided by meta-classification loss on the selected pseudo-labeled target samples. Therefore, this network can fit a weighting function according to the cross-user WHAR task at hand, which is superior to existing sample differentiation rules fixed for special scenarios. Extensive experiments on three public WHAR datasets demonstrate that SWL-Adapt achieves the state-of-the-art performance on the cross-user WHAR task, outperforming the best baseline by an average of 3.1% and 5.3% in accuracy and macro F1 score, respectively.