Provable Self-Play Algorithms for Competitive Reinforcement Learning

Provable Self-Play Algorithms for Competitive Reinforcement Learning


Self-play, where the algorithm learns by playing against itself without requiring any direct supervision, has become the new weapon in modern Reinforcement Learning (RL) for achieving superhuman performance in practice. However, the majority of exisiting theory in reinforcement learning only applies to the setting where the agent plays against a fixed environment. It remains largely open whether self-play algorithms can be provably effective, especially when it is necessary to manage the exploration/exploitation tradeoff.

We study self-play in competitive reinforcement learning under the setting of Markov games, a generalization of Markov decision processes to the two-player case. We introduce a self-play algorithm—Value Iteration with Upper/Lower Confidence Bound (VI-ULCB), and show that it achieves regret after playing steps of the game. The regret is measured by the agent’s performance against a fully adversarial opponent who can exploit the agent’s strategy at any step. We also introduce an explore-then-exploit style algorithm, which achieves a slightly worse regret of , but is guaranteed to run in polynomial time even in the worst case. To the best of our knowledge, our work presents the first line of provably sample-efficient self-play algorithms for competitive reinforcement learning.

1 Introduction

This paper studies competitive reinforcement learning (competitive RL), that is, reinforcement learning with two or more agents taking actions simultaneously, but each maximizing their own reward. Competitive RL is a major branch of the more general setting of multi-agent reinforcement learning (MARL), with the specification that the agents have conflicting rewards (so that they essentially compete with each other) yet can be trained in a centralized fashion (i.e. each agent has access to the other agents’ policies)   (Crandall and Goodrich, 2005).

There are substantial recent progresses in competitive RL, in particular in solving hard multi-player games such as GO (Silver et al., 2017), Starcraft (Vinyals et al., 2019), and Dota 2 (OpenAI, 2018). A key highlight in their approaches is the successful use of self-play for achieving super-human performance in absence of human knowledge or expert opponents. These self-play algorithms are able to learn a good policy for all players from scratch through repeatedly playing the current policies against each other and performing policy updates using these self-played game trajectories. The empirical success of self-play has challenged the conventional wisdom that expert opponents are necessary for achieving good performance, and calls for a better theoretical understanding.

In this paper, we take initial steps towards understanding the effectiveness of self-play algorithms in competitive RL from a theoretical perspective. We focus on the special case of two-player zero-sum Markov games (Shapley, 1953; Littman, 1994), a generaliztion of Markov Decision Processes (MDPs) to the two-player setting. In a Markov game, the two players share states, play actions simultaneously, and observe the same reward. However, one player aims to maximize the return while the other aims to minimize it. This setting covers the majority of two-player games including GO (there is a single reward of at the end of the game indicating which player has won), and also generalizes zero-sum matrix games (von Neumann, 1928)—an important game-theoretic problem—into the multi-step (RL) case.

More concretely, the goal of this paper is to design low-regret algorithms for solving episodic two-player Markov games in the general setting (Kearns and Singh, 2002), that is, the algorithm is allowed to play the game for a fixed amount of episodes using arbitrary policies, and its performance is measured in terms of the regret. We consider a strong notion of regret for two-player zero-sum games, where the performance of the deployed policies in each episode is measured against the best response for that policy, which can be different in differnet episodes. Such a regret bound measures the algorithm’s ability in managing the exploration and exploitation tradeoff against fully adaptive opponents, and can directly translate to other types of guarantees such as the PAC sample complexity bound.

Settings Algorithm Regret PAC Runtime
General Markov Game VI-ULCB (Theorem 3.2) PPAD-complete
VI-explore (Theorem 4) Polynomial
Mirror Descent () (Rakhlin and Sridharan, 2013)
Turn-Based Markov Game VI-ULCB (Corollary 3.3)
Mirror Descent () (Theorem 5)
Both Lower Bound (Corollary 5) -
Table 1: Regret and PAC guarantees of the Algorithms in this paper for zero-sum Markov games.

Our contribution

This paper introduces the first line of provably sample-efficient self-play algorithms for zero-sum Markov game under no restrictive assumptions. Concretely,

  • We introduce the first self-play algorithm with regret for zero-sum Markov games. More specifically, it achieves regret in the general case, where is the length of the game, is the number of states, are the number of actions for each player, and is the total number of steps played. In special case of turn-based games, it achieves regret with guaranteed polynomial runtime.

  • We also introduce an explore-then-exploit style algorithm. It has guaranteed polynomial runtime in the general setting of zero-sum Markov games, with a slightly worse regret.

  • We raise the open question about the optimal dependency of the regret on . We provide a lower bound , and show that the lower bound can be achieved in simple case of two-step turn-based games by a mirror descent style algorithm.

Above results are summarized in Table 1.

1.1 Related Work

There is a fast-growing body of work on multi-agent reinforcement learning (MARL). Many of them achieve striking empirical performance, or attack MARL in the cooperative setting, where agents are optimizing for a shared or similar reward. We refer the readers to several recent surveys for these results (see e.g. Buşoniu et al., 2010; Nguyen et al., 2018; OroojlooyJadid and Hajinezhad, 2019; Zhang et al., 2019). In the rest of this section we focus on theoretical results related to competitive RL.

Markov games

Markov games (or stochastic games) is proposed as a mathematical model for compeitive RL back in the early 1950s (Shapley, 1953). There is a long line of classical work since then on solving this problem (see e.g. Littman, 1994, 2001; Hu and Wellman, 2003; Hansen et al., 2013). They design algorithms, possibly with runtime guarantees, to find optimal policies in Markov games when both the transition matrix and reward are known, or in the asymptotic setting where number of data goes to infinity. These results do not directly apply to the non-asymptotic setting where the transition and reward are unknown and only a limited amount of data are available for estimating them.

A few recent work tackles self-play algorithms for Markov games in the non-asymptotic setting, working under either structural assumptions about the game or stronger sampling oracles. Wei et al. (2017) propose an upper confidence algorithm for stochastic games and prove that a self-play style algorithm finds -optimal policies in samples. Jia et al. (2019); Sidford et al. (2019) study turn-based stochastic games—a special case of general Markov games, and propose algorithms with near-optimal sample complexity. However, both lines of work make strong assumptions—on either the structure of Markov games or how we access data—that are not always true in practice. Specifically, Wei et al. (2017) assumes no matter what strategy one agent sticks to, the other agent can always reach all states by playing a certain policy, and Jia et al. (2019); Sidford et al. (2019) assume access to simulators (or generative models) which enable the agent to directly sample transition and reward information for any state-action pair. These assumptions greatly alleviate the challenge in exploration. In contrast, our results apply to general Markov games without further structural assumptions, and our algorithms have built-in mechanisms for solving the challenge in the exploration-exploitation tradeoff.

Finally, we note that classical R-max algorithm (Brafman and Tennenholtz, 2002) does not make restrictive assumptions. It also has provable guarantees even when playing against the adversarial opponent in Markov game. However, the theoretical guarantee in Brafman and Tennenholtz (2002) is weaker than the standard regret, and does not directly imply any self-play algorithm with regret bound in our setting (See Section E for more details).

Adversarial MDP

Another line of related work focuses on provable algorithms against adversarial opponents in MDP. Most work in this line considers the setting with adversarial rewards (see e.g. Zimin and Neu, 2013; Rosenberg and Mansour, 2019; Jin et al., 2019). These results do not direcly imply provable self-play algorithms in our setting, because the adversarial opponent in Markov games can affect both the reward and the transition. There exist a few works that tackle both adversarial transition functions and adversarial rewards (Yu and Mannor, 2009; Cheung et al., 2019; Lykouris et al., 2019). In particular, Lykouris et al. (2019) considers a stochastic problem with episodes arbitrarily corrupted and obtain regret. When applying these results to Markov games with an adversarial opponent, can be without further assumptions, which makes the bound vacuous.

Single-agent RL

There is an extensive body of research on the sample efficiency of reinforcement learning in the single agent setting (see e.g. Jaksch et al., 2010; Osband et al., 2014; Azar et al., 2017; Dann et al., 2017; Strehl et al., 2006; Jin et al., 2018), which are studied under the model of Markov decision process—a special case of Markov games. For the tabular episodic setting with nonstationary dynamics and no simulators, the best regrets achieved by existing model-based and model-free algorithms are (Azar et al., 2017) and (Jin et al., 2018), respectively, where is the number of states, is the number of actions, is the length of each episode, and is the total number of steps played. Both of them (nearly) match the minimax lower bound  (Jaksch et al., 2010; Osband and Van Roy, 2016; Jin et al., 2018).

2 Preliminaries

In this paper, we consider zero-sum Markov Games (MG) (Shapley, 1953; Littman, 1994), which also known as stochastic games in the literature. Zero-sum Markov games are generalization of standard Markov Decision Processes (MDP) into the two-player setting, in which the max-player seeks to maximize the total return and the min-player seeks to minimize the total return.

Formally, we consider tabular episodic zero-sum Markov games of the form , where

  • is the number of steps in each episode.

  • , and is the set of states at step , with .

  • , and is the set of actions of the max-player at step , with .

  • , and is the set of actions of the min-player at step , with .

  • is a collection of transition matrices, so that gives the distribution over states if action pair is taken for state at step .

  • is a collection of reward functions, and is the deterministic reward function at step . Note that we are assuming that rewards are in for normalization. 1

In each episode of this MG, an initial state is picked arbitrarily by an adversary. Then, at each step , both players observe state , pick the action simultaneously, receive reward , and then transition to the next state . The episode ends when is reached.

Policy and value function

A policy of the max-player is a collection of functions , where is the probability simplex over action set . Similarly, a policy of the min-player is a collection of functions . We use the notation and to present the probability of taking action or for state at step under policy or respectively. We use to denote the value function at step under policy and , so that gives the expected cumulative rewards received under policy and , starting from , until the end of the episode:

We also define to denote -value function at step so that gives the cumulative rewards received under policy and , starting from , till the end of the episode:

For simplicity, we use notation of operator so that for any value function . By definition of value functions, for all , we have the Bellman equation


where we define for all

Best response and regret

We now define the notion of best response and review some basic properties of it (cf. (Filar and Vrieze, 2012)), which will motivate our definition of the regret in two-player Markov games. For any max-player strategy , there exists a best response of the min-player, which is a policy satisfying for any . For simplicity, we denote . By symmetry, we can define the best response of the max-player , and define . The value functions and satisfy the following Bellman optimality equations:


It is further known that there exist policies , that are optimal against the best responses of the opponent:

It is also known that, for any , the minimax theorem holds:

Therefore, the optimal strategies are also the Nash Equalibrium for the Markov game. Based on this, it is sensible to measure the suboptimality of any pair of policies using the gap between their performance and the performance of the optimal strategy when playing against the best responses respectively, i.e.,


We make this formal in the following definition of the regret. {definition}[Regret] For any algorithm that plays the Markov game for episodes with (potentially adversarial) starting state for each episode , the regret is defined as

where , denote the policies deployed by the algorithm in the -th episode.

We note that as a unique feature of self-play algorithms, the learner is playing against herself, and thus chooses strategies for both max-player and min-player at each episode.

2.1 Turn-based games

In zero-sum Markov games, each step involves the two players playing simultaneously and independently. It is a general framework, which constains a very important special case—turn-based games. (Shapley, 1953; Jia et al., 2019).

The main feature of a turn-based game is that only one player is taking actions in each step; in other words, the max and min player take turns to play the game. Formally, a turn-based game can be defined through a partition of steps into two sets and , where and denote the sets of steps the max-player and the min-player choose the actions respectively, whhich satisfies and . As a special example, GO is a turn-based game in which the two players play in alternate turns, i.e.

Mathematically, we can specialize general zero-sum Markov games to turn-based games by restricting for all , and for all , where and are special dummy actions. Consequently, in those steps, or has only a single action as its element, i.e. the corresponding player can not affect the game in those steps. A consequence of this specialization is that the Nash Equilibria for turn-based games are pure strategies (i.e. deterministic policies) (Shapley, 1953), similar as in one-player MDPs. This is not always true for general Markov games.

3 Main Results

In this section, we present our algorithm and main theorems. In particular, our algorithm is the first self-play algorithm that achieves regret in Markov Games. We describe the algorithm in Section 3.1, and present its theoretical guarantee for general Markov games in Section 3.2. In Section 3.3, we show that when specialized to turn-based games, the regret and runtime of our algorithm can be further improved.

3.1 Algorithm description

To solve zero-sum Markov games, the main idea is to extend the celebrated UCB (Upper Confidence Bounds) principle—an algorithmic principle that achieves provably efficient exploration in bandits (Auer et al., 2002) and single-agent RL (Azar et al., 2017; Jin et al., 2018)—to the two-player setting. Recall that in single-agent RL, the provably efficient UCBVI algorithm (Azar et al., 2017) proceeds as

Algorithm (UCBVI for single-player RL): Compute based on estimated transition and optimistic (upper) estimate of reward, then play one episode with the greedy policy with respect to .

Regret bounds for UCBVI is then established by showing and utilizing the fact that remains an optimistic (upper) estimate of the optimal throughout execution of the algorithm.

In zero-sum games, the two player have conflicting goals: the max-player seeks to maximize the return and the min-player seeks to minimize the return. Therefore, it seems natural here to maintain two sets of Q estimates, one upper bounding the true value and one lower bounding the true value, so that each player can play optimistically with respect to her own goal. We summarize this idea into the following proposal.

Proposal (Naive two-player extension of UCBVI): Compute based on estimated transition and {upper, lower} estimates of rewards, then play one episode where the max-player () is greedy with respect to and the min-player () is greedy with respect to .

However, the above proposal is not yet a well-defined algorithm: a greedy strategy with respect to requires the knowledge of how the other player chooses , and vice versa. Therefore, what we really want is not that “ is greedy with respect to ”, but rather that “ is greedy with respect to when the other player uses ”, and vice versa. In other words, we rather desire that are jointly greedy with respect to .

Our algorithm concretizes such joint greediness precisely, building on insights from one-step matrix games: we choose to be the Nash equilibrium for the general-sum game in which the payoff matrix for the max player is and for the min player is . In other words, both player have their own payoff matrix (and they are not equal), but they jointly determine their policies. Formally, we let be determined as

for all , where Nash_General_Sum is a subroutine that takes two matrices , and returns the Nash equilibrium for general sum game, which satisfies


Such an equilibrium is guaranteed to exist due to the seminal work of Nash (1951), and is computable by algorithms such as the Lemke-Howson algorithm (Lemke and Howson, 1964). With the Nash_General_Sum subroutine in hand, our algorithm can be briefly described as

Our Algorithm (VI-ULCB): Compute based on estimated transition and {upper, lower} estimates of rewards, along the way determining policies by running the Nash_General_Sum subroutine on . Play one episode according to .

The full algorithm is described in Algorithm 1.

1:  Initialize: for any , , , .
2:  for episode  do
3:     Receive .
4:     for step  do
5:        for  do
6:           ;
7:           if t=0 then
9:           else
12:        for  do
14:           .
15:           .
16:     for step  do
17:        Take action , .
18:        Observe reward and next state .
19:        .
21:        .
22:        .
Algorithm 1 Value Iteration with Upper-Lower Confidence Bound (VI-ULCB)

3.2 Guarantees for General Markov Games

We are now ready to present our main theorem.


[Regret bound for VI-ULCB] For zero-sum Markov games, Algorithm 1 (with choice of bonus for large absolute constant ) achieves regret

with probability at least , where . We defer the proof of Theorem 3.2 into Appendix A.1.

Optimism in the face of uncertainty and best response

An implication of Theorem 3.2 is that a low regret can be achieved via self-play, i.e. the algorithm plays with itself and does not need an expert as its opponent. This is intriguing because the regret is measured in terms of the suboptimality against the worst-case opponent:

(Note that this decomposition of the regret has a slightly different form from (5).) Therefore, Theorem 3.2 demonstrates that self-play can protect against fully adversarial opponent even when such a strong opponent is not explicitly available.

The key technical reason enabling such a guarantee is that our estimates are optimistic in the face of both the uncertainty of the game, as well as the best response from the opponent. More precisely, we show that the in Algorithm 1 satisfy with high probability

for all , where denote the running at the beginning of the -th episode (Lemma A.1). In constrast, such a guarantee (and consequently the regret bound) is not achievable if the upper and lower estimates are only guaranteed to {upper, lower} bound the values of the Nash equilibrium.

Translation to PAC bound

Our regret bound directly implies a PAC sample complexity bound for learning near-equilibrium policies, based on an online-to-batch conversion. We state this in the following Corollary, and defer the proof to Appendix A.2. {corollary}[PAC bound for VI-ULCB] Suppose the initial state of Markov game is fixed at , then there exists a pair of (randomized) policies derived through the VI-ULCB algorithm such that with probability at least (over the randomness in the trajectories) we have

as soon as the number of episodes , where , and the expectation is over the randomization in .

Runtime of Algorithm 1

Algorithm 1 involves the Nash_General_Sum subroutine for computing the Nash equilibrium of a general sum matrix game. However, it is known that the computational complexity for approximating2 such an equilibrium is PPAD-complete (Daskalakis, 2013), a complexity class conjectured to not enjoy polynomial or quasi-polynomial time algorithms. Therefore, Algorithm 1 is strictly speaking not a polynomial time algorithm, despite of being rather sample-efficient.

We note however that there exists practical implementations of the subroutine such as the Lemke-Howson algorithm (Lemke and Howson, 1964) that can usually find the solution efficiently. We will further revisit the computational issue in Section 4, in which we design a computationally efficient algorithm for zero-sum games with a slightly worse regret.

3.3 Guarantees for Turn-based Markov Games

We now instantiate Theorem 3.2 on turn-based games (introduced in Section 2.1), in which the same algorithm enjoys better regret guarantee and polynomial runtime. Recall that in turn-based games, for all , we have either or , therefore given and we have

and thus by Theorem 3.2 the regret of Algorithm 1 on turn-based games is bounded by .

Further, since either or , all the Nash_General_Sum subroutines reduce to vector games rather than matrix games, and can be trivially implemented in polynomial (indeed linear) time. Indeed, suppose the payoff matrices in (6) has dimensions , then Nash_General_Sum reduces to finding and such that

(the other side is trivialized as has only one choice), which is solved at where . The situation is similar if .

We summarize the above results into the following corollary. {corollary}[Regret bound for VI-ULCB on turn-based games] For turn-based zero-sum Markov games, Algorithm 1 has runtime and achieves regret bound with probability at least , where .

4 Computationally Efficient Algorithm

In this section, we show that the computational issue of Algorithm 1 is not intrinsic to the problem: there exists a sublinear regret algorithm for general zero-sum Markov games that has a guaranteed polynomial runtime, with regret scaling as , slightly worse than that of Algorithm 1. Therefore, computational efficiency can be achieved if one is willing to trade some statistical efficiency (sample complexity). For simplicity, we assume in this section that the initial state is fixed.

1:  .
2:   for any .
3:  for step  do
4:     for  do
5:        .
6:     for  do
8:        .
9:  for all remaining episodes do
10:     Play the game with policy .
Algorithm 2 Value Iteration after Exploration (VI-Explore)

Value Iteration after Exploration

At a high level, our algorithm follows an explore-then-exploit approach. We begin by running a (polynomial time) reward-free exploration procedure Reward_Free_Exploration on a small number of episodes, which queries the MDP and outputs an estimate . Then, we run value iteration on the empirial version of Markov game with transition and reward , which finds its Nash equilibrium . Finally, the algorithm simply plays the policy for the remaining episodes. The full algorithm is described in Algorithm 2 in the Appendix.

By “reward-free” exploration, we mean the procedure will not use any reward information to guide exploration. Instead, the procedure prioritize on visiting all possible states and gathering sufficient information about their transition and rewards, so that are close to in the sense that the Nash equilibria of and are close, where denotes the Markov game with transition and reward .

This goal can be achieved by the following algorithm. For any fixed state , we can create an artificial reward defined as and for any , and . Then, we can treat as a new action set for a single agent, and run any standard reinforcement learning algorithm with PAC or regret guarantees to find a near-optimal policy of MDP. It can be shown that the optimal policy for this MDP is the policy that maximize the probability to reach state . Therefore, by repeatedly playing , we can gather transition and reward information at state as well as we can. Finally, we repeat the routine above for all state . See Appendix B for more details.

In this paper, we adapt the sharp treatments in Jin et al. (2020) which studies reward-free exploration in the single-agent MDP setting, and provide following guarantees for the Reward_Free_Exploration procedure.


[PAC bound for VI-Explore] With probability at least , Reward_Free_Exploration runs for episodes with some large constant , and , and outputs such that the Nash equilibrium of MG satisfies

Importantly, such Nash equilibrium of MG can be computed by Value Iteration (VI) using and . VI only calls Nash_Zero_Sum subroutine, which takes a matrix and returns the Nash equilibrium for zero-sum game, which satisfies


This problem can by solved efficiently (in polynomial time) by many existing algorithms designed for convex-concave optimization (see, e.g. (Koller, 1994)), and does not suffer from the PPAD-completeness that Nash_General_Sum does.

The PAC bound in Theorem 4 can be easily converted into a regret bound, which is presented as follows.


[Polynomial time algorithm via explore-then-exploit] For zero-sum Markov games, with probability at least , Algorithm 2 runs in time, and achieves regret bound

where .

5 Towards the Optimal Regret

We investigate the tightness of our regret upper bounds in Theorem 3.2 and Corollary 3.3 through raising the question of optimal regret in two-player Markov games, and making initial progresses on it by providing lower bounds and new upper bounds in specific settings. Specifically, we ask an

Open question: What is the optimal regret for general Markov games (in terms of dependence on )?

It is known that the (tight) regret lower bound for single-player MDPs is  (Azar et al., 2017). By restricting two-player games to a single-player MDP (making the other player dummy), we immediately have


[Regret lower bound, corollary of Jaksch et al. (2010), Theorem 5] The regret3 for any algorithm on turn-based games (and thus also general zero-sum games) is lower bounded by . Comparing this lower bound with the upper bound in Theorem 3.2 ( regret for general games and regret for turn-based games), there are gaps in both the -dependence and the -dependence.

Matching the lower bound on short-horizon games

Towards closing the gap between lower and upper bounds, we develop alternative algorithms in the special case where each player only plays once, i.e. one-step general games with and two-step turn-based games. In these cases, we show that there exists mirror descent type algorithms that achieve an improved regret (and thus matching the lower bounds), provided that we consider a weaker notion of the regret defined as {definition}[Weak Regret] The weak regret for any algorithm that deploys policies in episode is defined as


The difference in the weak regret is that it uses fixed opponents—as opposed to adaptive opponents—for measuring the performance gap: the max is taken with respect to a fixed for all episodes , rather than a different for each episode. By definition, we have for any algorithm that .

With the definition of the weak regret in hand, we now present our results for one-step games. Their proofs can be found in Appendix C.


[Weak regret for one-step simultaneous game, adapted from Rakhlin and Sridharan (2013)] For one-step simultaneous games (), there exists a mirror descent type algorithm that achieves weak regret bound with high probability.


[Weak regret for two-step turn-based game] For one-step turn-based games (), there exists a mirror descent type algorithm that achieves weak regret bound with high probability.

Proof insights; bottleneck in multi-step case

The improved regret bounds in Theorem 5 and 5 are possible due to availability of unbiased estimates of counterfactual Q values, which in turn can be used in mirror descent type algorithms with guarantees. Such unbiased estimates are only achievable in one-step games as the two policies are “not intertwined” in a certain sense. In contrast, in multi-step games (where each player plays more than once), such unbiased estimates of counterfactual Q values are no longer available, and it is unclear how to construct a mirror descent algorithm there. We believe it would be an important open question to close the gap in multi-step games (as well as the gap between regret and weak regret) for a further understanding of exploration in games.

6 Conclusion

In this paper, we studied the sample complexity of finding the equilibrium policy in the setting of competitive reinforcement learning, i.e. zero-sum Markov games with two players. We designed a self-play algorithm for zero-sum games and showed that it can efficiently find the Nash equilibrium policy in the exploration setting through establishing a regret bound. Our algorithm—Value Iteration with Upper and Lower Confidence Bounds—builds on a principled extension of UCB/optimism into the two-player case by constructing upper and lower bounds on the value functions and iteratively solving general sum subgames.

Towards investigating the optimal runtime and sample complexity in two-player games, we provided accompanying results showing that (1) the computational efficiency of our algorithm can be improved by explore-then-exploit type algorithms, which has a slightly worse regret; (2) the state and action space dependence in the regret can be reduced in the special case of one-step games via alternative mirror descent type algorithms.

We believe this paper opens up many interesting directions for future work. For example, can we design a computationally efficient algorithms that achieves regret? What are the optimal dependence of the regret on in multi-step games? Also, the present results only work in tabular games, and it would be of interest to investigate if similar results can hold in presence of function approximation.


We thank Sham Kakade and Haipeng Luo for valuable discussions on the related work. We also thank the Simons Institute at Berkeley for hosting the authors and incubating our initial discussions.

Appendix A Proofs for Section 3

a.1 Proof of Theorem 3.2


To be clear from the context, we denote the upper bound and lower bound and computed at the -th episode as and , and policies computed and used at the -th episode as and .

Choice of bonus:

for sufficient large absolute constant .


[ULCB] With probability at least , we have following bounds for any :


By symmetry, we only need to prove the statement (9). For each fixed , we prove this by induction from to . For base case, we know at the -th step, .

Now, assume the left inequality in (9) holds for -th step, for the -th step, we first recall the updates for functions respectively:

In case of