Soft Preference Optimization: Aligning Language Models to Expert Distributions

We propose Soft Preference Optimization (SPO), a method for aligning generative models, such as Large Language Models (LLMs), with human preferences, without the need for a reward model. SPO optimizes model outputs directly over a preference dataset through a natural loss function that integrates preference loss with a regularization term across the model's entire output distribution rather than limiting it to the preference dataset. Although SPO does not require the assumption of an existing underlying reward model, we demonstrate that, under the Bradley-Terry (BT) model assumption, it converges to a softmax of scaled rewards, with the distribution's "softness" adjustable via the softmax exponent, an algorithm parameter. We showcase SPO's methodology, its theoretical foundation, and its comparative advantages in simplicity, computational efficiency, and alignment precision.

On the Importance of Uncertainty in Decision-Making with Large Language Models

We investigate the role of uncertainty in decision-making problems with natural language as input. For such tasks, using Large Language Models as agents has become the norm. However, none of the recent approaches employ any additional phase for estimating the uncertainty the agent has about the world during the decision-making task. We focus on a fundamental decision-making framework with natural language as input, which is the one of contextual bandits, where the context information consists of text. As a representative of the approaches with no uncertainty estimation, we consider an LLM bandit with a greedy policy, which picks the action corresponding to the largest predicted reward. We compare this baseline to LLM bandits that make active use of uncertainty estimation by integrating the uncertainty in a Thompson Sampling policy. We employ different techniques for uncertainty estimation, such as Laplace Approximation, Dropout, and Epinets. We empirically show on real-world data that the greedy policy performs worse than the Thompson Sampling policies. These findings suggest that, while overlooked in the LLM literature, uncertainty plays a fundamental role in bandit tasks with LLMs.

In-context Exploration-Exploitation for Reinforcement Learning

In-context learning is a promising approach for online policy learning of offline reinforcement learning (RL) methods, which can be achieved at inference time without gradient optimization. However, this method is hindered by significant computational costs resulting from the gathering of large training trajectory sets and the need to train large Transformer models. We address this challenge by introducing an In-context Exploration-Exploitation (ICEE) algorithm, designed to optimize the efficiency of in-context policy learning. Unlike existing models, ICEE performs an exploration-exploitation trade-off at inference time within a Transformer model, without the need for explicit Bayesian inference. Consequently, ICEE can solve Bayesian optimization problems as efficiently as Gaussian process biased methods do, but in significantly less time. Through experiments in grid world environments, we demonstrate that ICEE can learn to solve new RL tasks using only tens of episodes, marking a substantial improvement over the hundreds of episodes needed by the previous in-context learning method.

Auxiliary task discovery through generate-and-test

In this paper, we explore an approach to auxiliary task discovery in reinforcement learning based on ideas from representation learning. Auxiliary tasks tend to improve data efficiency by forcing the agent to learn auxiliary prediction and control objectives in addition to the main task of maximizing reward, and thus producing better representations. Typically these tasks are designed by people. Meta-learning offers a promising avenue for automatic task discovery; however, these methods are computationally expensive and challenging to tune in practice. In this paper, we explore a complementary approach to the auxiliary task discovery: continually generating new auxiliary tasks and preserving only those with high utility. We also introduce a new measure of auxiliary tasks usefulness based on how useful the features induced by them are for the main task. Our discovery algorithm significantly outperforms random tasks, hand-designed tasks, and learning without auxiliary tasks across a suite of environments.

Importance Sampling Placement in Off-Policy Temporal-Difference Methods

A central challenge to applying many off-policy reinforcement learning algorithms to real world problems is the variance introduced by importance sampling. In off-policy learning, the agent learns about a different policy than the one being executed. To account for the difference importance sampling ratios are often used, but can increase variance in the algorithms and reduce the rate of learning. Several variations of importance sampling have been proposed to reduce variance, with per-decision importance sampling being the most popular. However, the update rules for most off-policy algorithms in the literature depart from per-decision importance sampling in a subtle way; they correct the entire TD error instead of just the TD target. In this work, we show how this slight change can be interpreted as a control variate for the TD target, reducing variance and improving performance. Experiments over a wide range of algorithms show this subtle modification results in improved performance.

An Empirical Comparison of Off-policy Prediction Learning Algorithms in the Four Rooms Environment

Many off-policy prediction learning algorithms have been proposed in the past decade, but it remains unclear which algorithms learn faster than others. We empirically compare 11 off-policy prediction learning algorithms with linear function approximation on two small tasks: the Rooms task, and the High Variance Rooms task. The tasks are designed such that learning fast in them is challenging. In the Rooms task, the product of importance sampling ratios can be as large as $2^{14}$ and can sometimes be two. To control the high variance caused by the product of the importance sampling ratios, step size should be set small, which in turn slows down learning. The High Variance Rooms task is more extreme in that the product of the ratios can become as large as $2^{14}\times 25$. This paper builds upon the empirical study of off-policy prediction learning algorithms by Ghiassian and Sutton (2021). We consider the same set of algorithms as theirs and employ the same experimental methodology. The algorithms considered are: Off-policy TD($\lambda$), five Gradient-TD algorithms, two Emphatic-TD algorithms, Tree Backup($\lambda$), Vtrace($\lambda$), and ABTD($\zeta$). We found that the algorithms' performance is highly affected by the variance induced by the importance sampling ratios. The data shows that Tree Backup($\lambda$), Vtrace($\lambda$), and ABTD($\zeta$) are not affected by the high variance as much as other algorithms but they restrict the effective bootstrapping parameter in a way that is too limiting for tasks where high variance is not present. We observed that Emphatic TD($\lambda$) tends to have lower asymptotic error than other algorithms, but might learn more slowly in some cases. We suggest algorithms for practitioners based on their problem of interest, and suggest approaches that can be applied to specific algorithms that might result in substantially improved algorithms.

An Empirical Comparison of Off-policy Prediction Learning Algorithms on the Collision Task

Off-policy prediction -- learning the value function for one policy from data generated while following another policy -- is one of the most challenging subproblems in reinforcement learning. This paper presents empirical results with eleven prominent off-policy learning algorithms that use linear function approximation: five Gradient-TD methods, two Emphatic-TD methods, Off-policy TD($\lambda$), Vtrace, and versions of Tree Backup and ABQ modified to apply to a prediction setting. Our experiments used the Collision task, a small idealized off-policy problem analogous to that of an autonomous car trying to predict whether it will collide with an obstacle. We assessed the performance of the algorithms according to their learning rate, asymptotic error level, and sensitivity to step-size and bootstrapping parameters. By these measures, the eleven algorithms can be partially ordered on the Collision task. In the top tier, the two Emphatic-TD algorithms learned the fastest, reached the lowest errors, and were robust to parameter settings. In the middle tier, the five Gradient-TD algorithms and Off-policy TD($\lambda$) were more sensitive to the bootstrapping parameter. The bottom tier comprised Vtrace, Tree Backup, and ABQ; these algorithms were no faster and had higher asymptotic error than the others. Our results are definitive for this task, though of course experiments with more tasks are needed before an overall assessment of the algorithms' merits can be made.

A Generalized Projected Bellman Error for Off-policy Value Estimation in Reinforcement Learning

Many reinforcement learning algorithms rely on value estimation. However, the most widely used algorithms -- namely temporal difference algorithms -- can diverge under both off-policy sampling and nonlinear function approximation. Many algorithms have been developed for off-policy value estimation which are sound under linear function approximation, based on the linear mean-squared projected Bellman error (PBE). Extending these methods to the non-linear case has been largely unsuccessful. Recently, several methods have been introduced that approximate a different objective, called the mean-squared Bellman error (BE), which naturally facilities nonlinear approximation. In this work, we build on these insights and introduce a new generalized PBE, that extends the linear PBE to the nonlinear setting. We show how this generalized objective unifies previous work, including previous theory, and obtain new bounds for the value error of the solutions of the generalized objective. We derive an easy-to-use, but sound, algorithm to minimize the generalized objective which is more stable across runs, is less sensitive to hyperparameters, and performs favorably across four control domains with neural network function approximation.

Does Standard Backpropagation Forget Less Catastrophically Than Adam?

Catastrophic forgetting remains a severe hindrance to the broad application of artificial neural networks (ANNs), however, it continues to be a poorly understood phenomenon. Despite the extensive amount of work on catastrophic forgetting, we argue that it is still unclear how exactly the phenomenon should be quantified, and, moreover, to what degree all of the choices we make when designing learning systems affect the amount of catastrophic forgetting. We use various testbeds from the reinforcement learning and supervised learning literature to (1) provide evidence that the choice of which modern gradient-based optimization algorithm is used to train an ANN has a significant impact on the amount of catastrophic forgetting and show that--surprisingly--in many instances classical algorithms such as vanilla SGD experience less catastrophic forgetting than the more modern algorithms such as Adam. We empirically compare four different existing metrics for quantifying catastrophic forgetting and (2) show that the degree to which the learning systems experience catastrophic forgetting is sufficiently sensitive to the metric used that a change from one principled metric to another is enough to change the conclusions of a study dramatically. Our results suggest that a much more rigorous experimental methodology is required when looking at catastrophic forgetting. Based on our results, we recommend inter-task forgetting in supervised learning must be measured with both retention and relearning metrics concurrently, and intra-task forgetting in reinforcement learning must--at the very least--be measured with pairwise interference.

Gradient Temporal-Difference Learning with Regularized Corrections

It is still common to use Q-learning and temporal difference (TD) learning-even though they have divergence issues and sound Gradient TD alternatives exist-because divergence seems rare and they typically perform well. However, recent work with large neural network learning systems reveals that instability is more common than previously thought. Practitioners face a difficult dilemma: choose an easy to use and performant TD method, or a more complex algorithm that is more sound but harder to tune and all but unexplored with non-linear function approximation or control. In this paper, we introduce a new method called TD with Regularized Corrections (TDRC), that attempts to balance ease of use, soundness, and performance. It behaves as well as TD, when TD performs well, but is sound in cases where TD diverges. We empirically investigate TDRC across a range of problems, for both prediction and control, and for both linear and non-linear function approximation, and show, potentially for the first time, that gradient TD methods could be a better alternative to TD and Q-learning.