Nicholas E. Corrado

Nicholas E. Corrado, Yuxiao Qu, John U. Balis et al.

In offline reinforcement learning (RL), an RL agent learns to solve a task using only a fixed dataset of previously collected data. While offline RL has been successful in learning real-world robot control policies, it typically requires large amounts of expert-quality data to learn effective policies that generalize to out-of-distribution states. Unfortunately, such data is often difficult and expensive to acquire in real-world tasks. Several recent works have leveraged data augmentation (DA) to inexpensively generate additional data, but most DA works apply augmentations in a random fashion and ultimately produce highly suboptimal augmented experience. In this work, we propose Guided Data Augmentation (GuDA), a human-guided DA framework that generates expert-quality augmented data. The key insight behind GuDA is that while it may be difficult to demonstrate the sequence of actions required to produce expert data, a user can often easily characterize when an augmented trajectory segment represents progress toward task completion. Thus, a user can restrict the space of possible augmentations to automatically reject suboptimal augmented data. To extract a policy from GuDA, we use off-the-shelf offline reinforcement learning and behavior cloning algorithms. We evaluate GuDA on a physical robot soccer task as well as simulated D4RL navigation tasks, a simulated autonomous driving task, and a simulated soccer task. Empirically, GuDA enables learning given a small initial dataset of potentially suboptimal experience and outperforms a random DA strategy as well as a model-based DA strategy.

Nicholas E. Corrado, Josiah P. Hanna

Recently, data augmentation (DA) has emerged as a method for leveraging domain knowledge to inexpensively generate additional data in reinforcement learning (RL) tasks, often yielding substantial improvements in data efficiency. While prior work has demonstrated the utility of incorporating augmented data directly into model-free RL updates, it is not well-understood when a particular DA strategy will improve data efficiency. In this paper, we seek to identify general aspects of DA responsible for observed learning improvements. Our study focuses on sparse-reward tasks with dynamics-invariant data augmentation functions, serving as an initial step towards a more general understanding of DA and its integration into RL training. Experimentally, we isolate three relevant aspects of DA: state-action coverage, reward density, and the number of augmented transitions generated per update (the augmented replay ratio). From our experiments, we draw two conclusions: (1) increasing state-action coverage often has a much greater impact on data efficiency than increasing reward density, and (2) decreasing the augmented replay ratio substantially improves data efficiency. In fact, certain tasks in our empirical study are solvable only when the replay ratio is sufficiently low.

Nicholas E. Corrado, Josiah P. Hanna

On-policy reinforcement learning (RL) algorithms are typically characterized as algorithms that perform policy updates using i.i.d. trajectories collected by the agent's current policy. However, after observing only a finite number of trajectories, such on-policy sampling may produce data that fails to match the expected on-policy data distribution. This sampling error leads to high-variance gradient estimates that yield data-inefficient on-policy learning. Recent work in the policy evaluation setting has shown that non-i.i.d., off-policy sampling can produce data with lower sampling error w.r.t. the expected on-policy distribution than on-policy sampling can produce (Zhong et. al, 2022). Motivated by this observation, we introduce an adaptive, off-policy sampling method to reduce sampling error during on-policy policy gradient RL training. Our method, Proximal Robust On-Policy Sampling (PROPS), reduces sampling error by collecting data with a behavior policy that increases the probability of sampling actions that are under-sampled w.r.t. the current policy. We empirically evaluate PROPS on continuous-action MuJoCo benchmark tasks as well as discrete-action tasks and demonstrate that (1) PROPS decreases sampling error throughout training and (2) increases the data efficiency of on-policy policy gradient algorithms.

Nicholas E. Corrado, Wenyuan Huang, Josiah P. Hanna

Multi-task reinforcement learning (MTRL) aims to train a single agent to efficiently optimize performance across multiple tasks simultaneously. However, jointly optimizing all tasks often yields imbalanced learning: agents quickly solve easy tasks but learn slowly on harder ones. While prior work primarily attributes this imbalance to conflicting task gradients and proposes gradient manipulation or specialized architectures to address it, we instead focus on a distinct and under-explored challenge: imbalanced data allocation. Standard MTRL allocates an equal number of environment interactions to each task, which over-allocates data to easy tasks that require relatively few interactions to solve and under-allocates data to hard tasks that require substantially more experience to solve. To address this challenge, we introduce Distributionally Robust Adaptive Task Sampling (DRATS), an algorithm that adaptively prioritizes sampling tasks furthest from being solved. We derive DRATS by formalizing MTRL as a feasibility problem from which we derive a minimax objective for minimizing the worst-case return gap, the difference between a desired target return and the agent's return on a task. In benchmarks like MetaWorld-MT10 and MT50, DRATS improves data efficiency and increases worst-task performance compared to existing task sampling algorithms.

Josiah P. Hanna, Nicholas E. Corrado

Recent work on large language models has demonstrated the use of model-free reinforcement learning (RL) to train reasoning-like capabilities. The emergence of "thinking" through model-free RL is interesting as thinking actions neither produce reward nor change the external world state to one where the agent is more likely to get reward. This paper seeks to build a domain-independent understanding of when model-free RL will lead to such "thinking" as a strategy for reward maximization. To build this understanding, we first introduce a theoretical model which we call a thought Markov decision process (MDP). Thought MDPs minimally extend the classical MDP model to include an abstract notion of thought state and thought action. Using the thought MDP model, we prove the importance of policy initialization in determining whether or not thinking emerges and show formally that thought actions are equivalent to the agent choosing to perform a step of policy improvement before continuing to act. We then show that open-source LLMs satisfy the conditions that our theory predicts are necessary for model-free RL to produce thinking-like behavior. Finally, we hypothesize sufficient conditions that would enable thinking to be learned outside of language generation and introduce a toy domain where a combination of multi-task pre-training and designated thought actions enable more data-efficient RL compared to non-thinking agents.

Nicholas E. Corrado, Josiah P. Hanna

Independent on-policy policy gradient algorithms are widely used for multi-agent reinforcement learning (MARL) in cooperative and no-conflict games, but they are known to converge suboptimally when each agent's policy gradient points toward a suboptimal equilibrium. In this work, we identify a subtler failure mode that arises \textit{even when the expected policy gradients of all agents point toward an optimal solution.} After collecting a finite set of trajectories, stochasticity in independent action sampling can cause the joint data distribution to deviate from the expected joint on-policy distribution. This \textit{sampling error} w.r.t. the joint on-policy distribution produces inaccurate gradient estimates that can lead agents to converge suboptimally. In this paper, we investigate if joint sampling error can be reduced through coordinated action selection and whether doing so improves the reliability of policy gradient learning in MARL. Toward this end, we introduce an adaptive action sampling approach to reduce joint sampling error. Our method, Multi-Agent Proximal Robust On-Policy Sampling (MA-PROPS), uses a centralized behavior policy that we continually adapt to place larger probability on joint actions that are currently under-sampled w.r.t. the current joint policy. We empirically evaluate MA-PROPS in a diverse range of multi-agent games and demonstrate that (1) MA-PROPS reduces joint sampling error more efficiently than standard on-policy sampling and (2) improves the reliability of independent policy gradient algorithms, increasing the fraction of training runs that converge to an optimal joint policy.

Nicholas E. Corrado, Julian Katz-Samuels, Adithya Devraj et al.

When aligning large language models (LLMs), their performance on various tasks (such as being helpful, harmless, and honest) depends heavily on the composition of their training data. However, selecting a data mixture that achieves strong performance across all tasks is challenging. Existing approaches rely on large ablation studies, heuristics, or human intuition, but these can be prohibitively expensive and suboptimal. We study this problem in the setting of preference optimization via DPO and introduce AutoMixAlign (AMA), a theoretically-grounded algorithm that adaptively mixes datasets during training to balance performance across tasks. AMA first trains \textit{specialist models} for each task to determine losses that correspond to strong task performance. Then, it trains a generalist model using a novel minimax optimization that prioritizes tasks for which generalist model losses deviate most from specialist model losses. To optimize this problem, we propose two algorithms: (1) AMA-R, which adaptively reweights the objective to prioritize tasks, and (2) AMA-S, which adaptively adjusts how much data is sampled from each task to prioritize tasks. Both algorithms achieve a convergence rate of $O(1/\sqrt{T})$ in the convex case. AMA-R's convergence result follows from Sagawa et al. (2019), and we provide a convergence proof for AMA-S using online learning techniques such as EXP3. We evaluate AMA on several multitask alignment setups and find that AMA outperforms the standard alignment approach -- which simply optimizes the total loss across all tasks -- and also outperforms model merging methods.