We investigate safe multi-agent reinforcement learning, where agents seek to collectively maximize an aggregate sum of local objectives while satisfying their own safety constraints. The objective and constraints are described by {\it general utilities}, i.e., nonlinear functions of the long-term state-action occupancy measure, which encompass broader decision-making goals such as risk, exploration, or imitations. The exponential growth of the state-action space size with the number of agents presents challenges for global observability, further exacerbated by the global coupling arising from agents' safety constraints. To tackle this issue, we propose a primal-dual method utilizing shadow reward and $\kappa$-hop neighbor truncation under a form of correlation decay property, where $\kappa$ is the communication radius. In the exact setting, our algorithm converges to a first-order stationary point (FOSP) at the rate of $\mathcal{O}\left(T^{-2/3}\right)$. In the sample-based setting, we demonstrate that, with high probability, our algorithm requires $\widetilde{\mathcal{O}}\left(\epsilon^{-3.5}\right)$ samples to achieve an $\epsilon$-FOSP with an approximation error of $\mathcal{O}(\phi_0^{2\kappa})$, where $\phi_0\in (0,1)$. Finally, we demonstrate the effectiveness of our model through extensive numerical experiments.

Block-based visual programming environments play an increasingly important role in introducing computing concepts to K-12 students. In recent years, they have also gained popularity in neuro-symbolic AI, serving as a benchmark to evaluate general problem-solving and logical reasoning skills. The open-ended and conceptual nature of these visual programming tasks make them challenging, both for state-of-the-art AI agents as well as for novice programmers. A natural approach to providing assistance for problem-solving is breaking down a complex task into a progression of simpler subtasks; however, this is not trivial given that the solution codes are typically nested and have non-linear execution behavior. In this paper, we formalize the problem of synthesizing such a progression for a given reference block-based visual programming task. We propose a novel synthesis algorithm that generates a progression of subtasks that are high-quality, well-spaced in terms of their complexity, and solving this progression leads to solving the reference task. We show the utility of our synthesis algorithm in improving the efficacy of AI agents (in this case, neural program synthesizers) for solving tasks in the Karel programming environment (Pattis et al., 1995). Then, we conduct a user study to demonstrate that our synthesized progression of subtasks can assist a novice programmer in solving tasks in the Hour of Code: Maze Challenge (Code.org,

Worry over polarization has grown alongside the digital information consumption revolution. Where most scientific work considered user-generated and user-disseminated (i.e.,~Web 2.0) content as the culprit, the potential of purely increased access to information (or Web 1.0) has been largely overlooked. Here, we suggest that the shift to Web 1.0 alone could include a powerful mechanism of belief extremization. We study an empirically calibrated persuasive argument model with confirmation bias. We compare an offline setting -- in which a limited number of arguments is broadcast by traditional media -- with an online setting -- in which the agent can choose to watch contents within a very wide set of possibilities. In both cases, we assume that positive and negative arguments are balanced. The simulations show that the online setting leads to significantly more extreme opinions and amplifies initial prejudice.

A major challenge for deep reinforcement learning (DRL) agents is to collaborate with novel partners that were not encountered by them during the training phase. This is specifically worsened by an increased variance in action responses when the DRL agents collaborate with human partners due to the lack of consistency in human behaviors. Recent work have shown that training a single agent as the best response to a diverse population of training partners significantly increases an agent's robustness to novel partners. We further enhance the population-based training approach by introducing a Hierarchical Reinforcement Learning (HRL) based method for Human-AI Collaboration. Our agent is able to learn multiple best-response policies as its low-level policy while at the same time, it learns a high-level policy that acts as a manager which allows the agent to dynamically switch between the low-level best-response policies based on its current partner. We demonstrate that our method is able to dynamically adapt to novel partners of different play styles and skill levels in the 2-player collaborative Overcooked game environment. We also conducted a human study in the same environment to test the effectiveness of our method when partnering with real human subjects.

Redistribution of resources within a group as a method to reduce wealth inequality is a current area of debate. The evolutionary path to or away from wealth sharing is also a subject of active research. In order to investigate effects and evolution of wealth sharing, societies are simulated using a minimal model of a complex adapting system. These simulations demonstrate, for this artificial foraging society, that local sharing of resources reduces the economy's total wealth and increases wealth inequality. Evolutionary pressures strongly select against local sharing, whether globally or within a individual's clan, and select for asocial behaviors. By holding constant the gene for sharing resources among neighbors, from rich to poor, either with everyone or only within members of the same clan, social behavior is selected but total wealth and mean age are substantially reduced relative to non-sharing societies. The Gini coefficient is shown to be ineffective in measuring these changes in total wealth and wealth distributions, and, therefore, individual well-being. Only with sociality do strategies emerge that allow sharing clans to exclude or coexist with non-sharing clans. These strategies are based on spatial effects, emphasizing the importance of modeling movement mediated community assembly and coexistence as well as sociality.

Determining the complexity of election attack problems is a major research direction in the computational study of voting problems. The paper "Towards completing the puzzle: complexity of control by replacing, adding, and deleting candidates or voters" by Erd\'elyi et al. (JAAMAS 2021) provides a comprehensive study of the complexity of control problems. The sole open problem is constructive control by replacing voters for 2-Approval. We show that this case is in P, strengthening the recent RP (randomized polynomial-time) upper bound due to Fitzsimmons and Hemaspaandra (IJCAI 2022). We show this by transforming 2-Approval CCRV to weighted matching. We also use this approach to show that priced bribery for 2-Veto elections is in P. With this result, and the accompanying (unsurprising) result that priced bribery for 3-Veto elections is NP-complete, this settles the complexity for $k$-Approval and $k$-Veto standard control and bribery cases.

We present a data-driven optimization approach for robotic controlled deposition with a degradable tool. Existing methods make the assumption that the tool tip is not changing or is replaced frequently. Errors can accumulate over time as the tool wears away and this leads to poor performance in the case where the tool degradation is unaccounted for during deposition. In the proposed approach, we utilize visual and force feedback to update the unknown model parameters of our tool-tip. Subsequently, we solve a constrained finite time optimal control problem for tracking a reference deposition profile, where our robot plans with the learned tool degradation dynamics. We focus on a robotic drawing problem as an illustrative example. Using real-world experiments, we show that the error in target vs actual deposition decreases when learned degradation models are used in the control design.

Training multiple agents to coordinate is an important problem with applications in robotics, game theory, economics, and social sciences. However, most existing Multi-Agent Reinforcement Learning (MARL) methods are online and thus impractical for real-world applications in which collecting new interactions is costly or dangerous. While these algorithms should leverage offline data when available, doing so gives rise to the offline coordination problem. Specifically, we identify and formalize the strategy agreement (SA) and the strategy fine-tuning (SFT) challenges, two coordination issues at which current offline MARL algorithms fail. To address this setback, we propose a simple model-based approach that generates synthetic interaction data and enables agents to converge on a strategy while fine-tuning their policies accordingly. Our resulting method, Model-based Offline Multi-Agent Proximal Policy Optimization (MOMA-PPO), outperforms the prevalent learning methods in challenging offline multi-agent MuJoCo tasks even under severe partial observability and with learned world models.

Centralized Training with Decentralized Execution (CTDE) has recently emerged as a popular framework for cooperative Multi-Agent Reinforcement Learning (MARL), where agents can use additional global state information to guide training in a centralized way and make their own decisions only based on decentralized local policies. Despite the encouraging results achieved, CTDE makes an independence assumption on agent policies, which limits agents from adopting global cooperative information from each other during centralized training. Therefore, we argue that the existing CTDE framework cannot fully utilize global information for training, leading to an inefficient joint-policy exploration and even suboptimal results. In this paper, we introduce a novel Centralized Advising and Decentralized Pruning (CADP) framework for multi-agent reinforcement learning, that not only enables an efficacious message exchange among agents during training but also guarantees the independent policies for execution. Firstly, CADP endows agents the explicit communication channel to seek and take advice from different agents for more centralized training. To further ensure the decentralized execution, we propose a smooth model pruning mechanism to progressively constrain the agent communication into a closed one without degradation in agent cooperation capability. Empirical evaluations on StarCraft II micromanagement challenge and Google Research Football benchmarks and and across different MARL backbones demonstrate that the proposed framework achieves superior performance compared with the state-of-the-art counterparts. Our code is available at https://github.com/zyh1999/CADP.

Most existing works consider direct perturbations of victim's state/action or the underlying transition dynamics to show vulnerability of reinforcement learning agents under adversarial attacks. However, such direct manipulation may not always be feasible in practice. In this paper, we consider another common and realistic attack setup: in a multi-agent RL setting with well-trained agents, during deployment time, the victim agent $\nu$ is exploited by an attacker who controls another agent $\alpha$ to act adversarially against the victim using an \textit{adversarial policy}. Prior attack models under such setup do not consider that the attacker can confront resistance and thus can only take partial control of the agent $\alpha$, as well as introducing perceivable ``abnormal'' behaviors that are easily detectable. A provable defense against these adversarial policies is also lacking. To resolve these issues, we introduce a more general attack formulation that models to what extent the adversary is able to control the agent to produce the adversarial policy. Based on such a generalized attack framework, the attacker can also regulate the state distribution shift caused by the attack through an attack budget, and thus produce stealthy adversarial policies that can exploit the victim agent. Furthermore, we provide the first provably robust defenses with convergence guarantee to the most robust victim policy via adversarial training with timescale separation, in sharp contrast to adversarial training in supervised learning which may only provide {\it empirical} defenses.