A centerpiece of the ever-popular reinforcement learning from human feedback (RLHF) approach to fine-tuning autoregressive language models is the explicit training of a reward model to emulate human feedback, distinct from the language model itself. This reward model is then coupled with policy-gradient methods to dramatically improve the alignment between language model outputs and desired responses. In this work, we adopt a novel perspective wherein a pre-trained language model is itself simultaneously a policy, reward function, and transition function. An immediate consequence of this is that reward learning and language model fine-tuning can be performed jointly and directly, without requiring any further downstream policy optimization. While this perspective does indeed break the traditional agent-environment interface, we nevertheless maintain that there can be enormous statistical benefits afforded by bringing to bear traditional algorithmic concepts from reinforcement learning. Our experiments demonstrate one concrete instance of this through efficient exploration based on the representation and resolution of epistemic uncertainty. In order to illustrate these ideas in a transparent manner, we restrict attention to a simple didactic data generating process and leave for future work extension to systems of practical scale.

Multi-robot navigation is the task of finding trajectories for a team of robotic agents to reach their destinations as quickly as possible without collisions. In this work, we introduce a new problem: fair-delay multi-robot navigation, which aims not only to enable such efficient, safe travels but also to equalize the travel delays among agents in terms of actual trajectories as compared to the best possible trajectories. The learning of a navigation policy to achieve this objective requires resolving a nontrivial credit assignment problem with robotic agents having continuous action spaces. Hence, we developed a new algorithm called Navigation with Counterfactual Fairness Filter (NCF2). With NCF2, each agent performs counterfactual inference on whether it can advance toward its goal or should stay still to let other agents go. Doing so allows us to effectively address the aforementioned credit assignment problem and improve fairness regarding travel delays while maintaining high efficiency and safety. Our extensive experimental results in several challenging multi-robot navigation environments demonstrate the greater effectiveness of NCF2 as compared to state-of-the-art fairness-aware multi-agent reinforcement learning methods. Our demo videos and code are available on the project webpage: https://omron-sinicx.github.io/ncf2/

In this paper, we propose a novel algorithm for energy-efficient, low-latency, accurate inference at the wireless edge, in the context of 6G networks endowed with reconfigurable intelligent surfaces (RISs). We consider a scenario where new data are continuously generated/collected by a set of devices and are handled through a dynamic queueing system. Building on the marriage between Lyapunov stochastic optimization and deep reinforcement learning (DRL), we devise a dynamic learning algorithm that jointly optimizes the data compression scheme, the allocation of radio resources (i.e., power, transmission precoding), the computation resources (i.e., CPU cycles), and the RIS reflectivity parameters (i.e., phase shifts), with the aim of performing energy-efficient edge classification with end-to-end (E2E) delay and inference accuracy constraints. The proposed strategy enables dynamic control of the system and of the wireless propagation environment, performing a low-complexity optimization on a per-slot basis while dealing with time-varying radio channels and task arrivals, whose statistics are unknown. Numerical results assess the performance of the proposed RIS-empowered edge inference strategy in terms of trade-off between energy, delay, and accuracy of a classification task.

Lifelong learning agents aim to learn multiple tasks sequentially over a lifetime. This involves the ability to exploit previous knowledge when learning new tasks and to avoid forgetting. Modulating masks, a specific type of parameter isolation approach, have recently shown promise in both supervised and reinforcement learning. While lifelong learning algorithms have been investigated mainly within a single-agent approach, a question remains on how multiple agents can share lifelong learning knowledge with each other. We show that the parameter isolation mechanism used by modulating masks is particularly suitable for exchanging knowledge among agents in a distributed and decentralized system of lifelong learners. The key idea is that the isolation of specific task knowledge to specific masks allows agents to transfer only specific knowledge on-demand, resulting in robust and effective distributed lifelong learning. We assume fully distributed and asynchronous scenarios with dynamic agent numbers and connectivity. An on-demand communication protocol ensures agents query their peers for specific masks to be transferred and integrated into their policies when facing each task. Experiments indicate that on-demand mask communication is an effective way to implement distributed lifelong reinforcement learning and provides a lifelong learning benefit with respect to distributed RL baselines such as DD-PPO, IMPALA, and PPO+EWC. The system is particularly robust to connection drops and demonstrates rapid learning due to knowledge exchange.

What are the computational foundations of social grouping? Traditional approaches to this question have focused on verbal reasoning or simple (low-dimensional) quantitative models. In the real world, however, social preferences emerge when high-dimensional learning systems (brains and bodies) interact with high-dimensional sensory inputs during an animal's embodied interactions with the world. A deep understanding of social grouping will therefore require embodied models that learn directly from sensory inputs using high-dimensional learning mechanisms. To this end, we built artificial neural networks (ANNs), embodied those ANNs in virtual fish bodies, and raised the artificial fish in virtual fish tanks that mimicked the rearing conditions of real fish. We then compared the social preferences that emerged in real fish versus artificial fish. We found that when artificial fish had two core learning mechanisms (reinforcement learning and curiosity-driven learning), artificial fish developed fish-like social preferences. Like real fish, the artificial fish spontaneously learned to prefer members of their own group over members of other groups. The artificial fish also spontaneously learned to self-segregate with their in-group, akin to self-segregation behavior seen in nature. Our results suggest that social grouping can emerge from three ingredients: (1) reinforcement learning, (2) intrinsic motivation, and (3) early social experiences with in-group members. This approach lays a foundation for reverse engineering animal-like social behavior with image-computable models, bridging the divide between high-dimensional sensory inputs and social preferences.

Recently, reward-conditioned reinforcement learning (RCRL) has gained popularity due to its simplicity, flexibility, and off-policy nature. However, we will show that current RCRL approaches are fundamentally limited and fail to address two critical challenges of RCRL -- improving generalization on high reward-to-go (RTG) inputs, and avoiding out-of-distribution (OOD) RTG queries during testing time. To address these challenges when training vanilla RCRL architectures, we propose Bayesian Reparameterized RCRL (BR-RCRL), a novel set of inductive biases for RCRL inspired by Bayes' theorem. BR-RCRL removes a core obstacle preventing vanilla RCRL from generalizing on high RTG inputs -- a tendency that the model treats different RTG inputs as independent values, which we term ``RTG Independence". BR-RCRL also allows us to design an accompanying adaptive inference method, which maximizes total returns while avoiding OOD queries that yield unpredictable behaviors in vanilla RCRL methods. We show that BR-RCRL achieves state-of-the-art performance on the Gym-Mujoco and Atari offline RL benchmarks, improving upon vanilla RCRL by up to 11%.

We propose MIMOC: Motion Imitation from Model-Based Optimal Control. MIMOC is a Reinforcement Learning (RL) controller that learns agile locomotion by imitating reference trajectories from model-based optimal control. MIMOC mitigates challenges faced by other motion imitation RL approaches because the references are dynamically consistent, require no motion retargeting, and include torque references. Hence, MIMOC does not require fine-tuning. MIMOC is also less sensitive to modeling and state estimation inaccuracies than model-based controllers. We validate MIMOC on the Mini-Cheetah in outdoor environments over a wide variety of challenging terrain, and on the MIT Humanoid in simulation. We show cases where MIMOC outperforms model-based optimal controllers, and show that imitating torque references improves the policy's performance.

The layout design of pipelines is a critical task in the construction industry. Currently, pipeline layout is designed manually by engineers, which is time-consuming and laborious. Automating and streamlining this process can reduce the burden on engineers and save time. In this paper, we propose a method for generating three-dimensional layout of pipelines based on deep reinforcement learning (DRL). Firstly, we abstract the geometric features of space to establish a training environment and define reward functions based on three constraints: pipeline length, elbow, and installation distance. Next, we collect data through interactions between the agent and the environment and train the DRL model. Finally, we use the well-trained DRL model to automatically design a single pipeline. Our results demonstrate that DRL models can complete the pipeline layout task in space in a much shorter time than traditional algorithms while ensuring high-quality layout outcomes.

Social dilemmas are situations where groups of individuals can benefit from mutual cooperation but conflicting interests impede them from doing so. This type of situations resembles many of humanity's most critical challenges, and discovering mechanisms that facilitate the emergence of cooperative behaviors is still an open problem. In this paper, we study the behavior of self-interested rational agents that learn world models in a multi-agent reinforcement learning (RL) setting and that coexist in environments where social dilemmas can arise. Our simulation results show that groups of agents endowed with world models outperform all the other tested ones when dealing with scenarios where social dilemmas can arise. We exploit the world model architecture to qualitatively assess the learnt dynamics and confirm that each agent's world model is capable to encode information of the behavior of the changing environment and the other agent's actions. This is the first work that shows that world models facilitate the emergence of complex coordinated behaviors that enable interacting agents to ``understand'' both environmental and social dynamics.

Exploration is a fundamental aspect of reinforcement learning (RL), and its effectiveness is a deciding factor in the performance of RL algorithms, especially when facing sparse extrinsic rewards. Recent studies have shown the effectiveness of encouraging exploration with intrinsic rewards estimated from novelties in observations. However, there is a gap between the novelty of an observation and an exploration, as both the stochasticity in the environment and the agent's behavior may affect the observation. To evaluate exploratory behaviors accurately, we propose DEIR, a novel method in which we theoretically derive an intrinsic reward with a conditional mutual information term that principally scales with the novelty contributed by agent explorations, and then implement the reward with a discriminative forward model. Extensive experiments on both standard and advanced exploration tasks in MiniGrid show that DEIR quickly learns a better policy than the baselines. Our evaluations on ProcGen demonstrate both the generalization capability and the general applicability of our intrinsic reward. Our source code is available at https://github.com/swan-utokyo/deir.