Deep reinforcement learning has a variety of different algorithms that solves many types of complex problems in various situations, one class of these algorithms is policy gradient (PG), which applies to a wide range of problems in both discrete and continuous action spaces, but applying it naively is inefficient, because of its poor sample complexity and high variance, which result in slower learning, to mitigate this we can use a baseline.
In this article, I will explain:
Let us dive in 🙂
Recall that the formula of policy gradient is
In practice, we approximate this expectation by taking N samples and average their values.
The cause of the high variance problem is the reward scale, we think of policy gradient as it increases the probability of taking good actions and decreases it for bad actions, but mostly this is not the case, imagine a situation where the “good” episode return was 10 and the “bad” one was 5, then both probabilities of the actions in those episodes will be increased, which is not what we want, this problem is what baselines can solve.
Mathematically, a baseline is a function when added to an expectation, does not change the expected value (or does not introduce bias), but at the same time, it can significantly affect the variance.
Following this definition, we want a baseline for the policy gradient that can reduce its high variance and does not change its direction, a natural thing to do is to take the actions that are better than average, increase their probability, and decrease the probability of the actions that are worse than average, this is implemented by calculating the average reward over the trajectory and subtract it from the reward at the current timestep, this kind of baselines is called the average reward baseline.
where b is the average reward over the sampled trajectories
Now, we will show how baselines do not change the expected value, and we can choose any baselines we want.
For notational convenience, we will write the policy gradient in the form
where
by the distributive property and linearity of expectation equation (1) can be expressed as a sum of two terms
Now, we will focus on the second term in equation (3), the goal is to show that it is equal to zero, which proves that adding a baseline will not change the expected value.
But before we dig in, I will remind you of identity from calculus.
after some rearrangement
Now, apply the definition of expectation to the second term of equation(3) and substitute the underlined term with the left-hand side of equation(5).
Since b is constant with respect to tau and by the linearity of the gradient operator, we can write the previous equation as
we know that π(τ ) is a probability distribution so its integral must be one, and the derivative of a constant is zero, using this result equation (1) will be
and this shows that baselines do not change the expected value and we can choose any baseline, but the question is how we choose a baseline that reduces the variance of the policy gradient effectively?
One way to answer this is to find the optimum baseline, which is the one that will minimize the variance of the policy gradient, the direct way to compute it is firstly by calculating the variance of the PG, take its derivative and set it to zero and solve the resulted equation for b, which is exactly what we will do next.
Recall that the variance of a random variable X is defined as
Applying this definition to the PG yields
Now unpack the definition of the PG from equation 1
Notice that the second term in the previous equation is the baselined PG squared and we know that it is the same as the regular PG, hence we can replace the second term with the normal PG, this will make the derivative calculation easier since the second term will not depend on b which makes its derivative zero.
After replacing the second term the equation is
Now we are ready to take the derivative with respect to b, I will assume that you are familiar with derivative calculations and the chain rule.
To reach to the optimum b set the derivative to zero
We can distribute the squared derivative and use the linearity of expectation to decompose the equation into two terms
Since b is a constant with respect to the expected value, we can pull it out of the expectation, after some the rearrangement the optimum baseline is
This baseline can be seen as the expected reward weighted by the PG squared, but usually, in practice, the type of baseline used is the previously mentioned average reward baseline in which we compute the average reward overall all trajectories and subtract it from the reward of the current trajectory.
In fact, there is a connection between the average reward baseline and the optimum one, you can read more details in section 3 of the paper by Lex and Nigel.
In this article we show how baselines can improve the variance of the PG without changing its expected value, we also derived the optimum baseline and introduced the average reward baseline.
Faisal Ahmed
My name is Faisal Ahmed, I graduated from the University of Khartoum, Sudan, with a Mechanical Engineering Degree, I worked as a Teaching Assistant and a Research Assistant, I am a machine learning enthusiast especially about deep reinforcement learning, I write articles about various
topics in ML and mathematics.
REINFORCE Algorithm: Taking baby steps in reinf...
Introduction to Gradient Descent Algorithm (alo...
What is Reinforcement Learning and How Does It ...
Nuts & Bolts of Reinforcement Learning: Mo...
Reinforcement Learning Techniques Based on Type...
Reinforcement Learning Guide: From Fundamentals...
A Hands-On Introduction to Deep Q-Learning usin...
Bellman Optimality Equation in Reinforcement Le...
Gradient Descent Algorithm: How Does it Work in...
Understanding Gradient Descent Algorithm
We use cookies essential for this site to function well. Please click to help us improve its usefulness with additional cookies. Learn about our use of cookies in our Privacy Policy & Cookies Policy.
Show details
This site uses cookies to ensure that you get the best experience possible. To learn more about how we use cookies, please refer to our Privacy Policy & Cookies Policy.
It is needed for personalizing the website.
Expiry: Session
Type: HTTP
This cookie is used to prevent Cross-site request forgery (often abbreviated as CSRF) attacks of the website
Expiry: Session
Type: HTTPS
Preserves the login/logout state of users across the whole site.
Expiry: Session
Type: HTTPS
Preserves users' states across page requests.
Expiry: Session
Type: HTTPS
Google One-Tap login adds this g_state cookie to set the user status on how they interact with the One-Tap modal.
Expiry: 365 days
Type: HTTP
Used by Microsoft Clarity, to store and track visits across websites.
Expiry: 1 Year
Type: HTTP
Used by Microsoft Clarity, Persists the Clarity User ID and preferences, unique to that site, on the browser. This ensures that behavior in subsequent visits to the same site will be attributed to the same user ID.
Expiry: 1 Year
Type: HTTP
Used by Microsoft Clarity, Connects multiple page views by a user into a single Clarity session recording.
Expiry: 1 Day
Type: HTTP
Collects user data is specifically adapted to the user or device. The user can also be followed outside of the loaded website, creating a picture of the visitor's behavior.
Expiry: 2 Years
Type: HTTP
Use to measure the use of the website for internal analytics
Expiry: 1 Years
Type: HTTP
The cookie is set by embedded Microsoft Clarity scripts. The purpose of this cookie is for heatmap and session recording.
Expiry: 1 Year
Type: HTTP
Collected user data is specifically adapted to the user or device. The user can also be followed outside of the loaded website, creating a picture of the visitor's behavior.
Expiry: 2 Months
Type: HTTP
This cookie is installed by Google Analytics. The cookie is used to store information of how visitors use a website and helps in creating an analytics report of how the website is doing. The data collected includes the number of visitors, the source where they have come from, and the pages visited in an anonymous form.
Expiry: 399 Days
Type: HTTP
Used by Google Analytics, to store and count pageviews.
Expiry: 399 Days
Type: HTTP
Used by Google Analytics to collect data on the number of times a user has visited the website as well as dates for the first and most recent visit.
Expiry: 1 Day
Type: HTTP
Used to send data to Google Analytics about the visitor's device and behavior. Tracks the visitor across devices and marketing channels.
Expiry: Session
Type: PIXEL
cookies ensure that requests within a browsing session are made by the user, and not by other sites.
Expiry: 6 Months
Type: HTTP
use the cookie when customers want to make a referral from their gmail contacts; it helps auth the gmail account.
Expiry: 2 Years
Type: HTTP
This cookie is set by DoubleClick (which is owned by Google) to determine if the website visitor's browser supports cookies.
Expiry: 1 Year
Type: HTTP
this is used to send push notification using webengage.
Expiry: 1 Year
Type: HTTP
used by webenage to track auth of webenagage.
Expiry: Session
Type: HTTP
Linkedin sets this cookie to registers statistical data on users' behavior on the website for internal analytics.
Expiry: 1 Day
Type: HTTP
Use to maintain an anonymous user session by the server.
Expiry: 1 Year
Type: HTTP
Used as part of the LinkedIn Remember Me feature and is set when a user clicks Remember Me on the device to make it easier for him or her to sign in to that device.
Expiry: 1 Year
Type: HTTP
Used to store information about the time a sync with the lms_analytics cookie took place for users in the Designated Countries.
Expiry: 6 Months
Type: HTTP
Used to store information about the time a sync with the AnalyticsSyncHistory cookie took place for users in the Designated Countries.
Expiry: 6 Months
Type: HTTP
Cookie used for Sign-in with Linkedin and/or to allow for the Linkedin follow feature.
Expiry: 6 Months
Type: HTTP
allow for the Linkedin follow feature.
Expiry: 1 Year
Type: HTTP
often used to identify you, including your name, interests, and previous activity.
Expiry: 2 Months
Type: HTTP
Tracks the time that the previous page took to load
Expiry: Session
Type: HTTP
Used to remember a user's language setting to ensure LinkedIn.com displays in the language selected by the user in their settings
Expiry: Session
Type: HTTP
Tracks percent of page viewed
Expiry: Session
Type: HTTP
Indicates the start of a session for Adobe Experience Cloud
Expiry: Session
Type: HTTP
Provides page name value (URL) for use by Adobe Analytics
Expiry: Session
Type: HTTP
Used to retain and fetch time since last visit in Adobe Analytics
Expiry: 6 Months
Type: HTTP
Remembers a user's display preference/theme setting
Expiry: 6 Months
Type: HTTP
Remembers which users have updated their display / theme preferences
Expiry: 6 Months
Type: HTTP
Used by Google Adsense, to store and track conversions.
Expiry: 3 Months
Type: HTTP
Save certain preferences, for example the number of search results per page or activation of the SafeSearch Filter. Adjusts the ads that appear in Google Search.
Expiry: 2 Years
Type: HTTP
Save certain preferences, for example the number of search results per page or activation of the SafeSearch Filter. Adjusts the ads that appear in Google Search.
Expiry: 2 Years
Type: HTTP
Save certain preferences, for example the number of search results per page or activation of the SafeSearch Filter. Adjusts the ads that appear in Google Search.
Expiry: 2 Years
Type: HTTP
Save certain preferences, for example the number of search results per page or activation of the SafeSearch Filter. Adjusts the ads that appear in Google Search.
Expiry: 2 Years
Type: HTTP
Save certain preferences, for example the number of search results per page or activation of the SafeSearch Filter. Adjusts the ads that appear in Google Search.
Expiry: 2 Years
Type: HTTP
Save certain preferences, for example the number of search results per page or activation of the SafeSearch Filter. Adjusts the ads that appear in Google Search.
Expiry: 2 Years
Type: HTTP
These cookies are used for the purpose of targeted advertising.
Expiry: 6 Hours
Type: HTTP
These cookies are used for the purpose of targeted advertising.
Expiry: 1 Month
Type: HTTP
These cookies are used to gather website statistics, and track conversion rates.
Expiry: 1 Month
Type: HTTP
Aggregate analysis of website visitors
Expiry: 6 Months
Type: HTTP
This cookie is set by Facebook to deliver advertisements when they are on Facebook or a digital platform powered by Facebook advertising after visiting this website.
Expiry: 4 Months
Type: HTTP
Contains a unique browser and user ID, used for targeted advertising.
Expiry: 2 Months
Type: HTTP
Used by LinkedIn to track the use of embedded services.
Expiry: 1 Year
Type: HTTP
Used by LinkedIn for tracking the use of embedded services.
Expiry: 1 Day
Type: HTTP
Used by LinkedIn to track the use of embedded services.
Expiry: 6 Months
Type: HTTP
Use these cookies to assign a unique ID when users visit a website.
Expiry: 6 Months
Type: HTTP
These cookies are set by LinkedIn for advertising purposes, including: tracking visitors so that more relevant ads can be presented, allowing users to use the 'Apply with LinkedIn' or the 'Sign-in with LinkedIn' functions, collecting information about how visitors use the site, etc.
Expiry: 6 Months
Type: HTTP
Used to make a probabilistic match of a user's identity outside the Designated Countries
Expiry: 90 Days
Type: HTTP
Used to collect information for analytics purposes.
Expiry: 1 year
Type: HTTP
Used to store session ID for a users session to ensure that clicks from adverts on the Bing search engine are verified for reporting purposes and for personalisation
Expiry: 1 Day
Type: HTTP
Cookie declaration last updated on 24/03/2023 by Analytics Vidhya.
Cookies are small text files that can be used by websites to make a user's experience more efficient. The law states that we can store cookies on your device if they are strictly necessary for the operation of this site. For all other types of cookies, we need your permission. This site uses different types of cookies. Some cookies are placed by third-party services that appear on our pages. Learn more about who we are, how you can contact us, and how we process personal data in our Privacy Policy.
Edit
Resend OTP
Resend OTP in 45s
Hi guys, just a quick unrelated question. Where can I subscribe to your awesome newsletter? I cant find the section for this on your website anymore.
Correct me if I am wrong, but you only demonstrated how a baseline doesn't impact the bias in the private case of a constant baseline that equals the average of all rewards. No? And then you talk about the advantage of baselines, all sorts of baselines. Can you prove the general case?