Implicit Policy for Reinforcement Learning

06/10/2018 ∙ by Yunhao Tang, et al. ∙ Columbia University 1

We introduce Implicit Policy, a general class of expressive policies that can flexibly represent complex action distributions in reinforcement learning, with efficient algorithms to compute entropy regularized policy gradients. We empirically show that, despite its simplicity in implementation, entropy regularization combined with a rich policy class can attain desirable properties displayed under maximum entropy reinforcement learning framework, such as robustness and multi-modality.



There are no comments yet.


page 8

page 20

This week in AI

Get the week's most popular data science and artificial intelligence research sent straight to your inbox every Saturday.

1 Introduction

Reinforcement Learning (RL) combined with deep neural networks have led to a wide range of successful applications, including the game of Go, robotics control and video game playing

[32, 30, 24]. During the training of deep RL agent, the injection of noise into the learning procedure can usually prevent the agent from premature convergence to bad locally optimal solutions, for example, by entropy regularization [30, 23] or by explicitly optimizing a maximum entropy objective [13, 25].

Though entropy regularization is much simpler to implement in practice, it greedily optimizes the policy entropy at each time step, without accounting for future effects. On the other hand, maximum entropy objective considers the entropy of the distribution over entire trajectories, and is more conducive to theoretical analysis [2]. Recently, [13, 14] also shows that optimizing the maximum entropy objective can lead to desirable properties such as robustness and multi-modal policy.

Can we preserve the simplicity of entropy regularization while attaining desirable properties under maximum entropy framework? To achieve this, a necessary condition is an expressive representation of policy. Though various flexible probabilistic models have been proposed in generative modeling [10, 37], such models are under-explored in policy based RL. To address such issues, we propose flexible policy classes and efficient algorithms to compute entropy regularized policy gradients.

In Section 3, we introduce Implicit Policy, a generic policy representation from which we derive two expressive policy classes, Normalizing Flows Policy (NFP) and more generally, Non-invertible Blackbox Policy (NBP). NFP provides a novel architecture that embeds state information into Normalizing Flows; NBP assumes little about policy architecture, yet we propose algorithms to efficiently compute entropy regularized policy gradients when the policy density is not accessible. In Section 4, we show that entropy regularization optimizes a lower bound of maximum entropy objective. In Section 5, we show that when combined with entropy regularization, expressive policies achieve competitive performance on benchmarks and leads to robust and multi-modal policies.

2 Preliminaries

2.1 Background

We consider the standard RL formalism consisting of an agent interacting with the environment. At time step , the agent is in state , takes action , receives instant reward and transitions to next state . Let be a policy. The objective of RL is to search for a policy which maximizes cumulative expected reward , where is a discount factor. The action value function of policy is defined as . In policy based RL, a policy is explicitly parameterized as with parameter , and the policy can be updated by policy gradients , where is the learning rate. So far, there are in general two ways to compute policy gradients for either on-policy or off-policy updates.

Score function gradient & Pathwise gradient.

Given a stochastic policy , the score function gradient for on-policy update is computed as as in [31, 30, 23, 35]. For off-policy update, it is necessary to introduce importance sampling weights to adjust the distribution difference between the behavior policy and current policy. Given a deterministic policy , the pathwise gradient for on-policy update is computed as . In practice, this gradient is often computed off-policy [33, 32], where the exact derivation comes from a modified off-policy objective [3].

Entropy Regularization.

For on-policy update, it is common to apply entropy regularization [38, 26, 23, 31]. Let be the entropy of policy at state . The entropy regularized update is


where is a regularization constant. By boosting policy entropy, this update can potentially prevent the policy from premature convergence to bad locally optimal solutions. In Section 3, we will introduce expressive policies that leverage both on-policy/off-policy updates, and algorithms to efficiently compute entropy regularized policy gradients.

Maximum Entropy RL.

In maximum entropy RL formulation, the objective is to maximize the cumulative reward and the policy entropy , where is a tradeoff constant. Note that differs from the update in (1) by an exchange of expectation and gradient. The intuition of

is to achieve high reward while being as random as possible over trajectories. Since there is no simple low variance gradient estimate for

, several previous works [31, 13, 25] have proposed to optimize primarily using off-policy value based algorithms.

2.2 Related Work

A large number of prior works have implemented policy gradient algorithms with entropy regularization [30, 31, 23, 26], which boost exploration by greedily maximizing policy entropy at each time step. In contrast to such greedy procedure, maximum entropy objective considers entropy over the entire policy trajectories [13, 25, 29]. Though entropy regularization is simpler to implement in practice, [12, 13] argues in favor of maximum entropy objective by showing that trained policies can be robust to noise, which is desirable for real life robotics tasks; and multi-modal, a potentially desired property for exploration and fine-tuning for downstream tasks. However, their training procedure is fairly complex, which consists of training a soft Q function by fixed point iteration and a neural sampler by Stein variational gradient [21]. We argue that properties as robustness and multi-modality are attainable through simple entropy regularized policy gradient algorithms combined with expressive policy representations.

Prior works have studied the property of maximum entropy objective [25, 39], entropy regularization [26] and their connections with variants of operators [2]. It is commonly believed that entropy regularization greedily maximizes local policy entropy and does not account for how a policy update impacts future states. In Section 4, we show that entropy regularized policy gradient update maximizes a lower bound of maximum entropy objective, given constraints on the differences between consecutive policy iterates. This partially justifies why simple entropy regularization combined with expressive policy classes can achieve competitive empirical performance in practice.

There is a number of prior works that discuss different policy architectures. The most common policy for continuous control is unimodal Gaussian [30, 31, 23]. [14] discusses mixtures of Gaussian, which can represent multi-modal policies but it is necessary to specify the number of modes in advance. [13] also represents a policy using implicit model, but the policy is trained to sample from the soft Q function instead of being trained directly. Recently, we find [11] also uses Normalizing Flows to represent policies, but their focus is learning an hierarchy and involves layers of pre-training. Contrary to early works, we propose to represent flexible policies using implicit models/Normalizing Flows and efficient algorithms to train the policy end-to-end.

Implicit models have been extensively studied in probabilistic inference and generative modeling [10, 17, 19, 37]

. Implicit models define distributions by transforming source noise via a forward pass of neural networks, which in general sacrifice tractable probability density for more expressive representation. Normalizing Flows are a special case of implicit models

[27, 5, 6], where transformations from source noise to output are invertible and allow for maximum likelihood inference. Borrowing inspirations from prior works, we introduce implicit models into policy representation and empirically show that such rich policy class entails multi-modal behavior during training. In [37], GAN [10] is used as an optimal density estimator for likelihood free inference. In our work, we apply similar idea to compute entropy regularization when policy density is not available.

3 Implicit Policy for Reinforcement Learning

We assume the action space to be a compact subset of . Any sufficiently smooth stochastic policy can be represented as a blackbox with parameter that incorporates state information and independent source noise sampled from a simple distribution . In state , the action is sampled by a forward pass in the blackbox.


For example, Gaussian policy is reduced to where is standard Gaussian [30]. In general, the distribution of is implicitly defined: for any set of , . Let be the density of this distribution222In future notations, when the context is clear, we use to denote both the density of the policy as well as the policy itself: for example, means sampling from the policy; means the log density of policy at in state .. We call such policy Implicit Policy as similar ideas have been previous explored in implicit generative modeling literature [10, 19, 37]. In the following, we derive two expressive stochastic policy classes following this blackbox formulation, and propose algorithms to efficiently compute entropy regularized policy gradients.

3.1 Normalizing Flows Policy (NFP)

We first construct a stochastic policy with Normalizing Flows. Normalizing Flows [27, 6] have been applied in variational inference and probabilistic modeling to represent complex distributions. In general, consider transforming a source noise by a series of invertible nonlinear function each with parameter , to output a target sample ,


Let be the Jacobian matrix of , then the density of

is computed by chain rule,


For a general invertible transformation , computing is expensive. We follow the architecture of [5] to ensure that is computed in linear time. To combine state information, we embed state by another neural network with parameter

and output a state vector

with the same dimension as . We can then insert the state vector between any two layers of (3) to make the distribution conditional on state . In our implementation, we insert the state vector after the first transformation (we detail our architecture design in Appendix C).


Though the additive form of and may in theory limit the capacity of the model, in practice we find the resulting policy still very expressive. For simplicity, we denote the above transformation (5) as with parameter . It is obvious that is still invertible between and , which is critical for computing according to (4). Such representations build complex policy distributions with explicit probability density , and hence entail training using score function gradient estimators.

Since there is no analytic form for entropy, we use samples to estimate entropy by re-parameterization, . The gradient of entropy can be easily computed by a pathwise gradient and easily implemented using back-propagation .

On-policy algorithm for NFP.

Any on-policy policy optimization algorithms can be easily combined with NFP. Since NFP has explicit access to policy density, it allows for training using score function gradient estimators with efficient entropy regularization.

3.2 Non-invertible Blackbox Policy (NBP)

The forward pass in (2) transforms the simple noise distribution to complex action distribution through the blackbox . However, the mapping is in general non-invertible and we do not have access to the density . We derive a pathwise gradient for such cases and leave all the proof in Appendix A.

Theorem 3.1 (Stochastic Pathwise Gradient).

Given an implicit stochastic policy . Let be the implicitly defined policy. Then the pathwise policy gradient for the stochastic policy is


To compute the gradient of policy entropy for such general implicit policy, we propose to train an additional classifier

with parameter along with policy . The classifier is trained to minimize the following objective given a policy



is a uniform distribution over


is the sigmoid function. We have

creftype A.1 in Appendix A.2 to guarantee that the optimal solution of (7) provides an estimate of policy density, . As a result, we could evaluate the entropy by simple re-parametrization . Further, we can compute gradients of the policy entropy through the density estimate as shown by the following theorem.

Theorem 3.2 (Unbiased Entropy Gradient).

Let be the optimal solution from (7), where the policy is given by implicit policy . The gradient of entropy can be computed as


It is worth noting that to compute , simply plugging in to replace in the entropy definition does not work in general, since the optimal solution of (7) implicitly depends on . However, fortunately in this case the additional term vanishes. The above theorem guarantees that we could apply entropy regularization even when the policy density is not accessible.

Off-policy algorithm for NBP.

We develop an off-policy algorithm for NBP. The agent contains an implicit with parameter , a critic with parameter and a classifier with parameter . At each time step , we sample action and save experience tuple to a replay buffer . During training, we sample a mini-batch of tuples from , update critic using TD learning, update policy using pathwise gradient (6) and update classifier by gradient descent on (7). We also maintain target networks with parameter to stabilize learning [24, 32]. The pseudocode is listed in Appendix D.

4 Entropy Regularization and Maximum Entropy RL

Though policy gradient algorithms with entropy regularization are easy to implement in practice, they are harder to analyze due to the lack of a global objective. Now we show that entropy regularization maximizes a lower bound of maximum entropy objective when consecutive policy iterates are close.

At each iteration of entropy regularized policy gradient algorithm, the policy parameter is updated as in (1). Following similar ideas in [15, 30], we now interpret such update as maximizing a linearized surrogate objective in the neighborhood of the previous policy iterate . The surrogate objective is


The first-order Taylor expansion of (9) centering at gives a linearized surrogate objective . Let , the entropy regularized update (1) is equivalent to solving the following optimization problem then update according to ,

where is a positive constant depending on both the learning rate and the previous iterate , and can be recovered from (1). The next theorem shows that by constraining the KL divergence of consecutive policy iterates, the surrogate objective (9) forms a non-trivial lower bound of maximum entropy objective,

Theorem 4.1 (Lower Bound).

If , then


By optimizing at each iteration, entropy regularized policy gradient algorithms maximize a lower bound of . This implies that though entropy regularization is a greedier procedure than optimizing maximum entropy objective, it accounts for certain effects that the maximum entropy objective is designed to capture. Nevertheless, the optimal solutions of both optimization procedures are different. Previous works [26, 13] have shown that the optimal solutions of both procedures are energy based policies, with energy functions being fixed points of Boltzmann operator and Mellowmax operator respectively [2]

. In Appendix B, we show that Boltzmann operator interpolates between Bellman operator and Mellowmax operator, which asserts that entropy regularization is greedier than optimizing

, yet it still maintains uncertainties in the policy updates.

Though maximum entropy objective accounts for long term effects of policy entropy updates and is more conducive to analysis [2], it is hard to implement a simple yet scalable procedure to optimize the objective [13, 14, 2]. Entropy regularization, on the other hand, is simple to implement in both on-policy and off-policy setting. In experiments, we will show that entropy regularized policy gradients combined with expressive policies achieve competitive performance in multiple aspects.

5 Experiments

Our experiments aim to answer the following questions: (1) Will expressive policy be hard to train, does implicit policy provide competitive performance on benchmark tasks? (2) Are implicit policies robust to noises on locomotion tasks? (3) Does implicit policy entropy regularization entail multi-modal policies as displayed under maximum entropy framework [13]?

To answer (1), we evaluate both NFP and NBP agent on benchmark continuous control tasks in MuJoCo [36] and compare with baselines. To answer (2), we compare NFP with unimodal Gaussian policy on locomotion tasks with additive observational noises. To answer (3), we illustrate the multi-modal capacity of both policy representations on specially designed tasks illustrated below, and compare with baselines. In all experiments, for NFP, we implement with standard PPO for on-policy update to approximately enforce the KL constraint (10) as in [31]; for NBP, we implement the off-policy algorithm developed in Section 3. In Appendix C and F, we detail hyper-parameter settings in the experiments and provide a small ablation study.

5.1 Locomotion Tasks

Benchmark tasks.

One potential disadvantage of expressive policies compared to simple policies (like unimodal Gaussian) is that they pose a more serious statistical challenge due to a larger number of parameters. To see if implicit policy suffers from such problems, we evaluate NFP and NBP on MuJoCo benchmark tasks. For each task, we train for a prescribed number of time steps, then report the results averaged over 5 random seeds. We compare the results with baseline algorithms, such as DDPG [32], SQL [13], TRPO [30] and PPO [31], where baseline TPRO and PPO use unimodal Gaussian policies. As can be seen from Table 1, both NFP and NBP achieve competitive performances on benchmark tasks: they outperform DDPG, SQL and TRPO on most tasks. However, baseline PPO tends to come on top on most tasks. Interestingly on HalfCheetah, baseline PPO gets stuck on a locally optimal gait, which NFP improves upon by a large margin.


Table 1: A comparison of implicit policy optimization with baseline algorithms on MuJoCo benchmark tasks. For each task, we show the average rewards achieved after training the agent for a fixed number of time steps. The results for NFP and NBP are averaged over 5 random seeds. The results for DDPG, SQL and TRPO are approximated based on the figures in [14], PPO is from OpenAI baseline implementation [4]. We highlight the top two algorithms for each task in bold font. Both TRPO and PPO use unimodal Gaussian policies.

Robustness to Noisy Observations.

We add independent Gaussian noise to each component of the observations to make the original tasks partially observable. Since PPO with unimodal Gaussian achieves leading performance on noise-free locomotion tasks across on-policy baselines (A2C [23], TRPO [30]) as shown in [31] and Appendix E.1, we compare NFP only with PPO with unimodal Gaussian on such noisy locomotion tasks. In Figure 1, we show the learning curves of both agents, where on many tasks NFP learns significantly faster than unimodal Gaussian. Why complex policies may add to robustness? We propose that since these control tasks are known to be solved by multiple separate modes of policy [22], observational noises potentially blur these modes and make it harder for a unimodal Gaussian policy to learn any single mode (e.g. unimodal Gaussian puts probability mass between two neighboring modes [18]). On the contrary, NFP can still navigate a more complex reward landscape thanks to a potentially multi-modal policy distribution and learn effectively. We leave a more detailed study of robustness, multi-modality and complex reward landscape as interesting future work.

(a) Hopper
(b) Walker
(c) Reacher
(d) Swimmer
(e) HalfCheetah
(f) Ant
(g) MountainCar
(h) Pendulum
Figure 1: Noisy Observations: learning curves on noisy locomotion tasks. For each task, the observation is added a Gaussian noise component-wise. Each curve is averaged over 4 random seeds. Red is NFP and blue is unimodal Gaussian, both implemented with PPO. NFP beats Gaussian on most tasks.

5.2 Multi-modal policy

Gaussian Bandits.

Though factorized unimodal policies suffice for most benchmark tasks, below we motivate the importance of a flexible policy by a simple example: Gaussian bandits. Consider a two dimensional bandit . The reward of action is for a positive definite matrix . The optimal policy for maximum entropy objective is , i.e. a Gaussian policy with covariance matrix . We compare NFP with PPO with factorized Gaussian. As illustrated in Figure 2(a), NFP can approximate the optimal Gaussian policy pretty closely while the factorized Gaussian cannot capture the high correlation between the two action components.

Navigating 2D Multi-goal.

We motivate the strength of implicit policy to represent multi-modal policy by Multi-goal environment [13]. The agent has 2D coordinates as states and 2D forces as actions . A ball is randomly initialized near the origin and the goal is to push the ball to reach one of the four goal positions plotted as red dots in Figure 2(b). While a unimodal policy can only deterministically commit the agent to one of the four goals, a multi-modal policy obtained by NBP can stochastically commit the agent to multiple goals. On the right of Figure 2(b) we also show sampled actions and contours of Q value functions at various states: NBP learns a very flexible policy with different number of modes in different states.

Learning a Bimodal Reacher.

For a more realistic example, consider learning a bimodal policy for reaching one of two targets (Figure 3(a)). The agent has the physical coordinates of the reaching arms as states and applies torques to the joints as actions . The objective is to move the reacher head to be close to one of the targets. As illustrated by trajectories in Figure 2(c), while a unimodal Gaussian policy can only deterministically reach one target (red curves), a NFP agent can capture both modes by stochastically reaching one of the two targets (blue curves).

(a) Gaussian Bandit
(b) 2D Multi-goal
(c) Bimodal Reacher
Figure 2: (a): Illustration of Gaussian bandits. The and

axes are actions. Green dots are actions from the optimal policy, a Gaussian distribution with covariance structure illustrated by the contours. Red dots and blue dots are actions sampled from a learned factorized Gaussian and NFP. NFP captures the covariance of the optimal policy while factorized Gaussian cannot. (b): Illustration of 2D multi-goal environment. Left: trajectories generated by trained NBP agent (solid blue curves). The

and axes are coordinates of the agent (state). The agent is initialized randomly near the origin. The goals are red dots, and instant rewards are proportional to the agent’s minimum distance to one of the four goals. Right: predicted Q value contours by the critic (light blue: low value, light green: high value and actions sampled from the policy (blue dots) at three selected states. The NFP policy has different number of modes at different states. (c): Trajectories of the reacher head by NFP (blue curves) and unimodal Gaussian policies (red curves) for the bimodal reacher. Yellow dots are locations of the two targets, and the green dot is the starting location of the reacher.

Fine-tuning for downstream tasks.

A recent paradigm for RL is to pre-train an agent to perform a conceptually high-level task, which may accelerate fine-tuning the agent to perform more specific tasks [13]. We consider pre-training a quadrupedal robot (Figure 3(b)) to run fast, then fine-tune the robot to run fast in a particular direction [13] as illustrated in Figure 3(c), where we set walls to limit the directions in which to run. Wide and Narrow Hallways tasks differ by the distance of the opposing walls. If an algorithm does not inject enough diversity during pre-training, it will commit the agent to prematurely run in a particular direction, which is bad for fine-tuning. We compare the pre-training capacity of DDPG [20], SQL [13] and NBP. As shown in Figure 3(d), after pre-training, NBP agent manages to run in multiple directions, while DDPG agent runs in a single direction due to a deterministic policy (Appendix E.2). In Table 2, we compare the cumulative rewards of agents after fine-tuning on downstream tasks with different pre-training as initializations. In both tasks, we find NBP to outperform DDPG, SQL and random initialization (no pre-training) by statistically significant margins, potentially because NBP agent learns a high-level running gait that is more conducive to fine-tuning. Interestingly, in Narrow Hallway, randomly initialized agent performs better than DDPG pre-training, which is probably because running fast in Narrow Hallway requires running in a very narrow direction, and DDPG pre-trained agent needs to first unlearn the overtly specialized running gait acquired from pre-training. In Wide Hallway, randomly initialized agent easily gets stuck in a locally optimal gait (running between two opposing walls) while pre-training in general helps avoid such problem.

(a) Reacher
(b) Ant
(c) Wide Hallway
(d) Ant Running
Figure 3: Illustration of locomotion tasks: (a) Bimodal Reacher. Train a reacher to reach one of two targets. Green boxes are targets. (b) Ant-Running. Train a quadrupedal robot to run fast. The instant reward is the robot’s center of mass velocity; (c) Ant-Hallway. Train a quadrupedal robot to run fast under the physical constraints of walls, the instant reward is the same as in (b). Narrow and Wide Hallway tasks differ by the distance between the opposing walls; (d) Trajectories by NBP agent in Ant-Running. The agent learns to run in multiple directions.
Tasks Random init DDPG init SQL init NBP init
Wide Hallway
Narrow Hallway

Table 2: A comparison of downstream fine-tuning under different initializations. For each task, we show the cumulative rewards after pre-training for steps and fine-tuning for steps. The rewards are shown in the form , all results are averaged over 5 seeds. Random init means the agent is trained from scratch.

Combining multiple modes by Imitation Learning.

We propose another paradigm that can be of practical interest. In general, learning a multi-modal policy from scratch is hard for complex tasks since it requires good exploration and an algorithm to learn multi-modal distributions [13], which is itself a hard inference problem [10]. A big advantage of policy based algorithm over value based algorithm [13]

is that the policy can be easily combined with imitation learning. We could decompose a complex task into several simpler tasks, each representing a simple mode of behavior easily learned by a RL agent, then combine them into a single agent using imitation learning or inverse RL

[1, 8, 28].

We illustrate with a stochastic Swimmer example (see Appendix E.3). Consider training a Swimmer to move fast either forward or backward. The aggregate behavior has two modes and it is easy to solve each single mode. We train two separate Swimmers to move forward/backward and generate expert trajectories using the trained agents. We then train a NBP / NFP agent using GAN [10] / maximum likelihood estimation to combine both modes. Training with the same algorithms, a unimodal policy either commits to only one mode or learns a policy that puts large probability mass between the two modes [18, 10], which greatly deviates from the expert policy. On the contrary, expressive policies can more flexibly incorporate multiple modes into a single agent.

6 Conclusion

We have proposed Implicit Policy, a rich class of policy that can represent complex action distributions. We have derived efficient algorithms to compute entropy regularized policy gradients for generic implicit policies. Importantly, we have also showed that entropy regularization maximizes a lower bound of maximum entropy objective, which implies that in practice entropy regularization rich policy class can lead to desired properties of maximum entropy RL. We have empirically showed that implicit policy achieves competitive performance on benchmark tasks, is more robust to observational noise, and can flexibly represent multi-modal distributions.


This research was supported by an Amazon Research Award (2017) and AWS cloud credits. The authors would like to thank Jalaj Bhandari for helpful discussions, and Sergey Levine for helpful comments on early stage experiments of the paper.


Appendix A Proof of Theorems

a.1 Stochastic Pathwise Gradient

See 3.1


We follow closely the derivation of deterministic policy gradient [33]. We assume that all conditions are satisfied to exchange expectations and gradients when necessary. Let denote the implicit policy . Let be the value function and action value function under such stochastic policy. We introduce as the probability of transitioning from to in steps under policy . Overloading the notation a bit, is the probability of in one step by taking action (i.e., ). We have

In the above derivation, we have used the Fubini theorem to interchange integral (expectation) and gradients. We can iterate the above derivation and have the following

With the above, we derive the pathwise policy gradient as follows

where is the discounted state visitation probability under policy . Writing the whole integral as an expectation over states, the policy gradient is

which is equivalent to in (6) in creftype 3.1. ∎

We can recover the result for deterministic policy gradient by using a degenerate functional form , i.e. with a deterministic function to compute actions.

a.2 Unbiased Entropy Gradient

Lemma A.1 (Optimal Classifier as Density Estimator).

Assume is expressive enough to represent any classifier (for example is a deep neural net). Assume to be bounded and let be uniform distribution over . Let be the optimizer to the optimization problem in (7). Then and is the volume of .


Observe that (7) is a binary classification problem with data from against . The optimal classifier of the problem produces the density ratio of these two distributions. See for example [10] for a detailed proof. ∎

See 3.2


Let be the density of implicit policy . The entropy is computed as follows

Computing its gradient


In the second line we highlight the fact that the expectation depends on parameter both implicitly through the density and through the sample . After decomposing the gradient using chain rule, we find that the first term vanishes, leaving the result shown in the theorem. ∎

a.3 Lower Bound

We recall that given a policy , the standard RL objective is . In maximum entropy formulation, the maximum entropy objective is


where is a regularization constant and is the entropy of policy at . We construct a surrogate objective based on another policy as follows


The following proof highly mimics the proof in [30]. We have the following definition for coupling two policies

Definition A.1 (coupled).

Two policies are coupled if for any .

Lemma A.2.

Given are coupled, then


Let denote the number of times that for , i.e. the number of times that disagree before time . We can decompose the expectations as follows

Note that implies for all hence

The definition of coupling implies , and so . Now we note that

Combining previous observations, we have proved the lemma. ∎

Note that if we take , then the surrogate objective in (9) is equivalent to defined in (13). With creftype A.2, we prove the following theorem.

See 4.1


We first show the result for general policies and with . As a result, [30] shows that are coupled. Recall the maximum entropy objective defined in (12) and surrogate objective in (13), take the difference of two objectives

Now observe that by taking , the above inequality implies the theorem. ∎

In practice, coupling enforced by KL divergence is often relaxed [30, 31]. The theorem implies that, by constraining the KL divergence between consecutive policy iterates, the surrogate objective of entropy regularization maximizes a lower bound of maximum entropy objective.

Appendix B Operator view of Entropy Regularization and Maximum Entropy RL

Recall in standard RL formulation, the agent is in state , takes action , receives reward and transitions to . Let the discount factor . Assume that the reward is deterministic and the transitions are deterministic, i.e. , it is straightforward to extend the following to general stochastic transitions. For a given policy , define linear Bellman operator as

Any policy satisfies the linear Bellman equation . Define Bellman optimality operator (we will call it Bellman operator) as

Now we define Mellowmax operator [2, 13] with parameter as follows,

It can be shown that both and are contractive operator when . Let be the unique fixed point of , then is the action value function of the optimal policy . Let be the unique fixed point of