Value decomposition is widely used in cooperative multi-agent reinforcement learning, however, its implicit credit assignment mechanism is not yet fully understood due to black-box networks. In this work, we study an interpretable value decomposition framework via the family of generalized additive models. We present a novel method, named Neural Attention Additive Q-learning (N$\text{A}^\text{2}$Q), providing inherent intelligibility of collaboration behavior. N$\text{A}^\text{2}$Q can explicitly factorize the optimal joint policy induced by enriching shape functions to model all possible coalitions of agents into individual policies. Moreover, we construct identity semantics to promote estimating credits together with the global state and individual value functions, where local semantic masks help us diagnose whether each agent captures relevant-task information. Extensive experiments show that N$\text{A}^\text{2}$Q consistently achieves superior performance compared to different state-of-the-art methods on all challenging tasks, while yielding human-like interpretability.

As machine learning (ML) algorithms are deployed for tasks with real-world social consequences (e.g., school admissions, loan approval, medical interventions, etc.), the possibility exists for runaway social inequalities (Crawford and Calo, 2016; Chaney et al., 2018; Fuster et al., 2018; Ensign et al., 2018). While "fairness" has become a salient ethical concern in contemporary research, the closed-loop dynamics of real-world systems comprising ML policies and populations that mutually adapt to each other (Figure 1 in the supplementary material) remain poorly understood. In this paper, our primary contribution is to consider the problem of long-term fairness, or algorithmic fairness in the context of a dynamically responsive population, as a reinforcement learning (RL) problem subject to constraint. In our formulation, the central learning task is to develop a policy that minimizes cumulative loss (e.g., financial risk, negative educational outcomes, misdiagnoses, etc.) incurred by an ML agent interacting with a human population up to a finite time horizon, subject to constraints on cumulative "violations of fairness", which we refer to in a single time step as disparity and cumulatively as distortion.

Combating an epidemic entails finding a plan that describes when and how to apply different interventions, such as mask-wearing mandates, vaccinations, school or workplace closures. An optimal plan will curb an epidemic with minimal loss of life, disease burden, and economic cost. Finding an optimal plan is an intractable computational problem in realistic settings. Policy-makers, however, would greatly benefit from tools that can efficiently search for plans that minimize disease and economic costs especially when considering multiple possible interventions over a continuous and complex action space given a continuous and equally complex state space. We formulate this problem as a Markov decision process. Our formulation is unique in its ability to represent multiple continuous interventions over any disease model defined by ordinary differential equations. We illustrate how to effectively apply state-of-the-art actor-critic reinforcement learning algorithms (PPO and SAC) to search for plans that minimize overall costs. We empirically evaluate the learning performance of these algorithms and compare their performance to hand-crafted baselines that mimic plans constructed by policy-makers. Our method outperforms baselines. Our work confirms the viability of a computational approach to support policy-makers

Humans naturally exploit haptic feedback during contact-rich tasks like loading a dishwasher or stocking a bookshelf. Current robotic systems focus on avoiding unexpected contact, often relying on strategically placed environment sensors. Recently, contact-exploiting manipulation policies have been trained in simulation and deployed on real robots. However, they require some form of real-world adaptation to bridge the sim-to-real gap, which might not be feasible in all scenarios. In this paper we train a contact-exploiting manipulation policy in simulation for the contact-rich household task of loading plates into a slotted holder, which transfers without any fine-tuning to the real robot. We investigate various factors necessary for this zero-shot transfer, like time delay modeling, memory representation, and domain randomization. Our policy transfers with minimal sim-to-real gap and significantly outperforms heuristic and learnt baselines. It also generalizes to plates of different sizes and weights. Demonstration videos and code are available at https://sites.google.com/view/compliant-object-insertion.

Robust Markov decision processes (RMDPs) provide a promising framework for computing reliable policies in the face of model errors. Many successful reinforcement learning algorithms build on variations of policy-gradient methods, but adapting these methods to RMDPs has been challenging. As a result, the applicability of RMDPs to large, practical domains remains limited. This paper proposes a new Double-Loop Robust Policy Gradient (DRPG), the first generic policy gradient method for RMDPs. In contrast with prior robust policy gradient algorithms, DRPG monotonically reduces approximation errors to guarantee convergence to a globally optimal policy in tabular RMDPs. We introduce a novel parametric transition kernel and solve the inner loop robust policy via a gradient-based method. Finally, our numerical results demonstrate the utility of our new algorithm and confirm its global convergence properties.

Covariate distribution shifts and adversarial perturbations present robustness challenges to the conventional statistical learning framework: mild shifts in the test covariate distribution can significantly affect the performance of the statistical model learned based on the training distribution. The model performance typically deteriorates when extrapolation happens: namely, covariates shift to a region where the training distribution is scarce, and naturally, the learned model has little information. For robustness and regularization considerations, adversarial perturbation techniques are proposed as a remedy; however, careful study needs to be carried out about what extrapolation region adversarial covariate shift will focus on, given a learned model. This paper precisely characterizes the extrapolation region, examining both regression and classification in an infinite-dimensional setting. We study the implications of adversarial covariate shifts to subsequent learning of the equilibrium -- the Bayes optimal model -- in a sequential game framework. We exploit the dynamics of the adversarial learning game and reveal the curious effects of the covariate shift to equilibrium learning and experimental design. In particular, we establish two directional convergence results that exhibit distinctive phenomena: (1) a blessing in regression, the adversarial covariate shifts in an exponential rate to an optimal experimental design for rapid subsequent learning, (2) a curse in classification, the adversarial covariate shifts in a subquadratic rate fast to the hardest experimental design trapping subsequent learning.

In this paper, we present DevFormer, a novel transformer-based architecture for addressing the complex and computationally demanding problem of hardware design optimization. Despite the demonstrated efficacy of transformers in domains including natural language processing and computer vision, their use in hardware design has been limited by the scarcity of offline data. Our approach addresses this limitation by introducing strong inductive biases such as relative positional embeddings and action-permutation symmetricity that effectively capture the hardware context and enable efficient design optimization with limited offline data. We apply DevFoemer to the problem of decoupling capacitor placement and show that it outperforms state-of-the-art methods in both simulated and real hardware, leading to improved performances while reducing the number of components by more than $30\%$. Finally, we show that our approach achieves promising results in other offline contextual learning-based combinatorial optimization tasks.

It is often very challenging to manually design reward functions for complex, real-world tasks. To solve this, one can instead use reward learning to infer a reward function from data. However, there are often multiple reward functions that fit the data equally well, even in the infinite-data limit. This means that the reward function is only partially identifiable. In this work, we formally characterise the partial identifiability of the reward function given several popular reward learning data sources, including expert demonstrations and trajectory comparisons. We also analyse the impact of this partial identifiability for several downstream tasks, such as policy optimisation. We unify our results in a framework for comparing data sources and downstream tasks by their invariances, with implications for the design and selection of data sources for reward learning.

Linear dynamical systems that obey stochastic differential equations are canonical models. While optimal control of known systems has a rich literature, the problem is technically hard under model uncertainty and there are hardly any results. We initiate study of this problem and aim to learn (and simultaneously deploy) optimal actions for minimizing a quadratic cost function. Indeed, this work is the first that comprehensively addresses the crucial challenge of balancing exploration versus exploitation in continuous-time systems. We present online policies that learn optimal actions fast by carefully randomizing the parameter estimates, and establish their performance guarantees: a regret bound that grows with square-root of time multiplied by the number of parameters. Implementation of the policy for a flight-control task demonstrates its efficacy. Further, we prove sharp stability results for inexact system dynamics and tightly specify the infinitesimal regret caused by sub-optimal actions. To obtain the results, we conduct a novel eigenvalue-sensitivity analysis for matrix perturbation, establish upper-bounds for comparative ratios of stochastic integrals, and introduce the new method of policy differentiation. Our analysis sheds light on fundamental challenges in continuous-time reinforcement learning and suggests a useful cornerstone for similar problems.

A combination of deep reinforcement learning and supervised learning is proposed for the problem of active sequential hypothesis testing in completely unknown environments. We make no assumptions about the prior probability, the action and observation sets, and the observation generating process. Our method can be used in any environment even if it has continuous observations or actions, and performs competitively and sometimes better than the Chernoff test, in both finite and infinite horizon problems, despite not having access to the environment dynamics.