Transformers Learn Temporal Difference Methods for In-Context Reinforcement Learning

In-context learning refers to the learning ability of a model during inference time without adapting its parameters. The input (i.e., prompt) to the model (e.g., transformers) consists of both a context (i.e., instance-label pairs) and a query instance. The model is then able to output a label for the query instance according to the context during inference. A possible explanation for in-context learning is that the forward pass of (linear) transformers implements iterations of gradient descent on the instance-label pairs in the context. In this paper, we prove by construction that transformers can also implement temporal difference (TD) learning in the forward pass, a phenomenon we refer to as in-context TD. We demonstrate the emergence of in-context TD after training the transformer with a multi-task TD algorithm, accompanied by theoretical analysis. Furthermore, we prove that transformers are expressive enough to implement many other policy evaluation algorithms in the forward pass, including residual gradient, TD with eligibility trace, and average-reward TD.

The ODE Method for Stochastic Approximation and Reinforcement Learning with Markovian Noise

Stochastic approximation is a class of algorithms that update a vector iteratively, incrementally, and stochastically, including, e.g., stochastic gradient descent and temporal difference learning. One fundamental challenge in analyzing a stochastic approximation algorithm is to establish its stability, i.e., to show that the stochastic vector iterates are bounded almost surely. In this paper, we extend the celebrated Borkar-Meyn theorem for stability from the Martingale difference noise setting to the Markovian noise setting, which greatly improves its applicability in reinforcement learning, especially in those off-policy reinforcement learning algorithms with linear function approximation and eligibility traces. Central to our analysis is the diminishing asymptotic rate of change of a few functions, which is implied by both a form of strong law of large numbers and a commonly used V4 Lyapunov drift condition and trivially holds if the Markov chain is finite and irreducible.

AlphaStar Unplugged: Large-Scale Offline Reinforcement Learning

StarCraft II is one of the most challenging simulated reinforcement learning environments; it is partially observable, stochastic, multi-agent, and mastering StarCraft II requires strategic planning over long time horizons with real-time low-level execution. It also has an active professional competitive scene. StarCraft II is uniquely suited for advancing offline RL algorithms, both because of its challenging nature and because Blizzard has released a massive dataset of millions of StarCraft II games played by human players. This paper leverages that and establishes a benchmark, called AlphaStar Unplugged, introducing unprecedented challenges for offline reinforcement learning. We define a dataset (a subset of Blizzard's release), tools standardizing an API for machine learning methods, and an evaluation protocol. We also present baseline agents, including behavior cloning, offline variants of actor-critic and MuZero. We improve the state of the art of agents using only offline data, and we achieve 90% win rate against previously published AlphaStar behavior cloning agent.

Direct Gradient Temporal Difference Learning

Off-policy learning enables a reinforcement learning (RL) agent to reason counterfactually about policies that are not executed and is one of the most important ideas in RL. It, however, can lead to instability when combined with function approximation and bootstrapping, two arguably indispensable ingredients for large-scale reinforcement learning. This is the notorious deadly triad. Gradient Temporal Difference (GTD) is one powerful tool to solve the deadly triad. Its success results from solving a doubling sampling issue indirectly with weight duplication or Fenchel duality. In this paper, we instead propose a direct method to solve the double sampling issue by simply using two samples in a Markovian data stream with an increasing gap. The resulting algorithm is as computationally efficient as GTD but gets rid of GTD's extra weights. The only price we pay is a logarithmically increasing memory as time progresses. We provide both asymptotic and finite sample analysis, where the convergence rate is on-par with the canonical on-policy temporal difference learning. Key to our analysis is a novel refined discretization of limiting ODEs.

Improving Monte Carlo Evaluation with Offline Data

Monte Carlo (MC) methods are the most widely used methods to estimate the performance of a policy. Given an interested policy, MC methods give estimates by repeatedly running this policy to collect samples and taking the average of the outcomes. Samples collected during this process are called online samples. To get an accurate estimate, MC methods consume massive online samples. When online samples are expensive, e.g., online recommendations and inventory management, we want to reduce the number of online samples while achieving the same estimate accuracy. To this end, we use off-policy MC methods that evaluate the interested policy by running a different policy called behavior policy. We design a tailored behavior policy such that the variance of the off-policy MC estimator is provably smaller than the ordinary MC estimator. Importantly, this tailored behavior policy can be efficiently learned from existing offline data, i,e., previously logged data, which are much cheaper than online samples. With reduced variance, our off-policy MC method requires fewer online samples to evaluate the performance of a policy compared with the ordinary MC method. Moreover, our off-policy MC estimator is always unbiased.

On the Chattering of SARSA with Linear Function Approximation

SARSA, a classical on-policy control algorithm for reinforcement learning, is known to chatter when combined with linear function approximation: SARSA does not diverge but oscillates in a bounded region. However, little is know about how fast SARSA converges to that region and how large the region is. In this paper, we make progress towards solving this open problem by showing the convergence rate of projected SARSA to a bounded region. Importantly, the region is much smaller than the ball used for projection provided that the the magnitude of the reward is not too large. Our analysis applies to expected SARSA as well as SARSA($\lambda$). Existing works regarding the convergence of linear SARSA to a fixed point all require the Lipschitz constant of SARSA's policy improvement operator to be sufficiently small; our analysis instead applies to arbitrary Lipschitz constants and thus characterizes the behavior of linear SARSA for a new regime.

Global Optimality and Finite Sample Analysis of Softmax Off-Policy Actor Critic under State Distribution Mismatch

In this paper, we establish the global optimality and convergence rate of an off-policy actor critic algorithm in the tabular setting without using density ratio to correct the discrepancy between the state distribution of the behavior policy and that of the target policy. Our work goes beyond existing works on the optimality of policy gradient methods in that existing works use the exact policy gradient for updating the policy parameters while we use an approximate and stochastic update step. Our update step is not a gradient update because we do not use a density ratio to correct the state distribution, which aligns well with what practitioners do. Our update is approximate because we use a learned critic instead of the true value function. Our update is stochastic because at each step the update is done for only the current state action pair. Moreover, we remove several restrictive assumptions from existing works in our analysis. Central to our work is the finite sample analysis of a generic stochastic approximation algorithm with time-inhomogeneous update operators on time-inhomogeneous Markov chains, based on its uniform contraction properties.

Truncated Emphatic Temporal Difference Methods for Prediction and Control

Emphatic Temporal Difference (TD) methods are a class of off-policy Reinforcement Learning (RL) methods involving the use of followon traces. Despite the theoretical success of emphatic TD methods in addressing the notorious deadly triad (Sutton and Barto, 2018) of off-policy RL, there are still three open problems. First, the motivation for emphatic TD methods proposed by Sutton et al. (2016) does not align with the convergence analysis of Yu (2015). Namely, a quantity used by Sutton et al. (2016) that is expected to be essential for the convergence of emphatic TD methods is not used in the actual convergence analysis of Yu (2015). Second, followon traces typically suffer from large variance, making them hard to use in practice. Third, despite the seminal work of Yu (2015) confirming the asymptotic convergence of some emphatic TD methods for prediction problems, there is still no finite sample analysis for any emphatic TD method for prediction, much less control. In this paper, we address those three open problems simultaneously via using truncated followon traces in emphatic TD methods. Unlike the original followon traces, which depend on all previous history, truncated followon traces depend on only finite history, reducing variance and enabling the finite sample analysis of our proposed emphatic TD methods for both prediction and control.

Learning Expected Emphatic Traces for Deep RL

Off-policy sampling and experience replay are key for improving sample efficiency and scaling model-free temporal difference learning methods. When combined with function approximation, such as neural networks, this combination is known as the deadly triad and is potentially unstable. Recently, it has been shown that stability and good performance at scale can be achieved by combining emphatic weightings and multi-step updates. This approach, however, is generally limited to sampling complete trajectories in order, to compute the required emphatic weighting. In this paper we investigate how to combine emphatic weightings with non-sequential, off-line data sampled from a replay buffer. We develop a multi-step emphatic weighting that can be combined with replay, and a time-reversed $n$-step TD learning algorithm to learn the required emphatic weighting. We show that these state weightings reduce variance compared with prior approaches, while providing convergence guarantees. We tested the approach at scale on Atari 2600 video games, and observed that the new X-ETD($n$) agent improved over baseline agents, highlighting both the scalability and broad applicability of our approach.

Breaking the Deadly Triad with a Target Network

The deadly triad refers to the instability of a reinforcement learning algorithm when it employs off-policy learning, function approximation, and bootstrapping simultaneously. In this paper, we investigate the target network as a tool for breaking the deadly triad, providing theoretical support for the conventional wisdom that a target network stabilizes training. We first propose and analyze a novel target network update rule which augments the commonly used Polyak-averaging style update with two projections. We then apply the target network and ridge regularization in several divergent algorithms and show their convergence to regularized TD fixed points. Those algorithms are off-policy with linear function approximation and bootstrapping, spanning both policy evaluation and control, as well as both discounted and average-reward settings. In particular, we provide the first convergent linear $Q$-learning algorithms under nonrestrictive and changing behavior policies without bi-level optimization.