Reinforcement Learning from Delayed Observations via World Models

In standard Reinforcement Learning settings, agents typically assume immediate feedback about the effects of their actions after taking them. However, in practice, this assumption may not hold true due to physical constraints and can significantly impact the performance of RL algorithms. In this paper, we focus on addressing observation delays in partially observable environments. We propose leveraging world models, which have shown success in integrating past observations and learning dynamics, to handle observation delays. By reducing delayed POMDPs to delayed MDPs with world models, our methods can effectively handle partial observability, where existing approaches achieve sub-optimal performance or even degrade quickly as observability decreases. Experiments suggest that one of our methods can outperform a naive model-based approach by up to %30. Moreover, we evaluate our methods on visual input based delayed environment, for the first time showcasing delay-aware reinforcement learning on visual observations.

Moonwalk: Inverse-Forward Differentiation

Backpropagation, while effective for gradient computation, falls short in addressing memory consumption, limiting scalability. This work explores forward-mode gradient computation as an alternative in invertible networks, showing its potential to reduce the memory footprint without substantial drawbacks. We introduce a novel technique based on a vector-inverse-Jacobian product that accelerates the computation of forward gradients while retaining the advantages of memory reduction and preserving the fidelity of true gradients. Our method, Moonwalk, has a time complexity linear in the depth of the network, unlike the quadratic time complexity of na\"ive forward, and empirically reduces computation time by several orders of magnitude without allocating more memory. We further accelerate Moonwalk by combining it with reverse-mode differentiation to achieve time complexity comparable with backpropagation while maintaining a much smaller memory footprint. Finally, we showcase the robustness of our method across several architecture choices. Moonwalk is the first forward-based method to compute true gradients in invertible networks in computation time comparable to backpropagation and using significantly less memory.

Skill Set Optimization: Reinforcing Language Model Behavior via Transferable Skills

Large language models (LLMs) have recently been used for sequential decision making in interactive environments. However, leveraging environment reward signals for continual LLM actor improvement is not straightforward. We propose Skill Set Optimization (SSO) for improving LLM actor performance through constructing and refining sets of transferable skills. SSO constructs skills by extracting common subtrajectories with high rewards and generating subgoals and instructions to represent each skill. These skills are provided to the LLM actor in-context to reinforce behaviors with high rewards. Then, SSO further refines the skill set by pruning skills that do not continue to result in high rewards. We evaluate our method in the classic videogame NetHack and the text environment ScienceWorld to demonstrate SSO's ability to optimize a set of skills and perform in-context policy improvement. SSO outperforms baselines by 40% in our custom NetHack task and outperforms the previous state-of-the-art in ScienceWorld by 35%.

Learning to Design Analog Circuits to Meet Threshold Specifications

Automated design of analog and radio-frequency circuits using supervised or reinforcement learning from simulation data has recently been studied as an alternative to manual expert design. It is straightforward for a design agent to learn an inverse function from desired performance metrics to circuit parameters. However, it is more common for a user to have threshold performance criteria rather than an exact target vector of feasible performance measures. In this work, we propose a method for generating from simulation data a dataset on which a system can be trained via supervised learning to design circuits to meet threshold specifications. We moreover perform the to-date most extensive evaluation of automated analog circuit design, including experimenting in a significantly more diverse set of circuits than in prior work, covering linear, nonlinear, and autonomous circuit configurations, and show that our method consistently reaches success rate better than 90% at 5% error margin, while also improving data efficiency by upward of an order of magnitude. A demo of this system is available at circuits.streamlit.app

Selective Perception: Optimizing State Descriptions with Reinforcement Learning for Language Model Actors

Large language models (LLMs) are being applied as actors for sequential decision making tasks in domains such as robotics and games, utilizing their general world knowledge and planning abilities. However, previous work does little to explore what environment state information is provided to LLM actors via language. Exhaustively describing high-dimensional states can impair performance and raise inference costs for LLM actors. Previous LLM actors avoid the issue by relying on hand-engineered, task-specific protocols to determine which features to communicate about a state and which to leave out. In this work, we propose Brief Language INputs for DEcision-making Responses (BLINDER), a method for automatically selecting concise state descriptions by learning a value function for task-conditioned state descriptions. We evaluate BLINDER on the challenging video game NetHack and a robotic manipulation task. Our method improves task success rate, reduces input size and compute costs, and generalizes between LLM actors.

Do Embodied Agents Dream of Pixelated Sheep?: Embodied Decision Making using Language Guided World Modelling

Reinforcement learning (RL) agents typically learn tabula rasa, without prior knowledge of the world, which makes learning complex tasks with sparse rewards difficult. If initialized with knowledge of high-level subgoals and transitions between subgoals, RL agents could utilize this Abstract World Model (AWM) for planning and exploration. We propose using few-shot large language models (LLMs) to hypothesize an AWM, that is tested and verified during exploration, to improve sample efficiency in embodied RL agents. Our DECKARD agent applies LLM-guided exploration to item crafting in Minecraft in two phases: (1) the Dream phase where the agent uses an LLM to decompose a task into a sequence of subgoals, the hypothesized AWM; and (2) the Wake phase where the agent learns a modular policy for each subgoal and verifies or corrects the hypothesized AWM on the basis of its experiences. Our method of hypothesizing an AWM with LLMs and then verifying the AWM based on agent experience not only increases sample efficiency over contemporary methods by an order of magnitude but is also robust to and corrects errors in the LLM, successfully blending noisy internet-scale information from LLMs with knowledge grounded in environment dynamics.

Reducing Variance in Temporal-Difference Value Estimation via Ensemble of Deep Networks

In temporal-difference reinforcement learning algorithms, variance in value estimation can cause instability and overestimation of the maximal target value. Many algorithms have been proposed to reduce overestimation, including several recent ensemble methods, however none have shown success in sample-efficient learning through addressing estimation variance as the root cause of overestimation. In this paper, we propose MeanQ, a simple ensemble method that estimates target values as ensemble means. Despite its simplicity, MeanQ shows remarkable sample efficiency in experiments on the Atari Learning Environment benchmark. Importantly, we find that an ensemble of size 5 sufficiently reduces estimation variance to obviate the lagging target network, eliminating it as a source of bias and further gaining sample efficiency. We justify intuitively and empirically the design choices in MeanQ, including the necessity of independent experience sampling. On a set of 26 benchmark Atari environments, MeanQ outperforms all tested baselines, including the best available baseline, SUNRISE, at 100K interaction steps in 16/26 environments, and by 68% on average. MeanQ also outperforms Rainbow DQN at 500K steps in 21/26 environments, and by 49% on average, and achieves average human-level performance using 200K ($\pm$100K) interaction steps. Our implementation is available at https://github.com/indylab/MeanQ.

Feasible Adversarial Robust Reinforcement Learning for Underspecified Environments

Robust reinforcement learning (RL) considers the problem of learning policies that perform well in the worst case among a set of possible environment parameter values. In real-world environments, choosing the set of possible values for robust RL can be a difficult task. When that set is specified too narrowly, the agent will be left vulnerable to reasonable parameter values unaccounted for. When specified too broadly, the agent will be too cautious. In this paper, we propose Feasible Adversarial Robust RL (FARR), a method for automatically determining the set of environment parameter values over which to be robust. FARR implicitly defines the set of feasible parameter values as those on which an agent could achieve a benchmark reward given enough training resources. By formulating this problem as a two-player zero-sum game, FARR jointly learns an adversarial distribution over parameter values with feasible support and a policy robust over this feasible parameter set. Using the PSRO algorithm to find an approximate Nash equilibrium in this FARR game, we show that an agent trained with FARR is more robust to feasible adversarial parameter selection than with existing minimax, domain-randomization, and regret objectives in a parameterized gridworld and three MuJoCo control environments.

Self-Play PSRO: Toward Optimal Populations in Two-Player Zero-Sum Games

In competitive two-agent environments, deep reinforcement learning (RL) methods based on the \emph{Double Oracle (DO)} algorithm, such as \emph{Policy Space Response Oracles (PSRO)} and \emph{Anytime PSRO (APSRO)}, iteratively add RL best response policies to a population. Eventually, an optimal mixture of these population policies will approximate a Nash equilibrium. However, these methods might need to add all deterministic policies before converging. In this work, we introduce \emph{Self-Play PSRO (SP-PSRO)}, a method that adds an approximately optimal stochastic policy to the population in each iteration. Instead of adding only deterministic best responses to the opponent's least exploitable population mixture, SP-PSRO also learns an approximately optimal stochastic policy and adds it to the population as well. As a result, SP-PSRO empirically tends to converge much faster than APSRO and in many games converges in just a few iterations.

Learning to Query Internet Text for Informing Reinforcement Learning Agents

Generalization to out of distribution tasks in reinforcement learning is a challenging problem. One successful approach improves generalization by conditioning policies on task or environment descriptions that provide information about the current transition or reward functions. Previously, these descriptions were often expressed as generated or crowd sourced text. In this work, we begin to tackle the problem of extracting useful information from natural language found in the wild (e.g. internet forums, documentation, and wikis). These natural, pre-existing sources are especially challenging, noisy, and large and present novel challenges compared to previous approaches. We propose to address these challenges by training reinforcement learning agents to learn to query these sources as a human would, and we experiment with how and when an agent should query. To address the \textit{how}, we demonstrate that pretrained QA models perform well at executing zero-shot queries in our target domain. Using information retrieved by a QA model, we train an agent to learn \textit{when} it should execute queries. We show that our method correctly learns to execute queries to maximize reward in a reinforcement learning setting.