I was reading a high level summary about Google's AlphaGo, and I came across the terms "policy network" and "value network". At a high level, I understand that the policy network is used to suggest moves and the value network is used to, "Reduce the depth of the search tree [and estimate] the winner in each position in place of searching all the way to the end of the game."

These two networks seem redundant to me. What is the policy network doing if it's not using the value network to prune its policies? It seems pretty clear that the value network is a deep learning neural network; is the policy network just a theoretical abstraction and not an actual neural network? The target variable for the value network seems to be win/loss. Is there a target variable for the policy network; if so, what is it? What is the policy network trying to optimize?

The full pdf of Google's paper, published in Nature, can be found here.

  • $\begingroup$ As a quick note: The target variable for the policy network is a probability distribution. From the paper: "This neural network (the value network) has a similar architecture to the policy network, but outputs a single prediction instead of a probability distribution" $\endgroup$
    – Zimano
    Mar 2, 2018 at 18:21
  • 3
    $\begingroup$ If anyone else stumbles upon this old question, like me, you'll be pleased to know that AlphaGo's successor, "AlphaGo Zero", as well as its successor "AlphaZero" do indeed get rid of the redundancy by only training ONE network whose results serve the functions of both the policy network (gives a probability distribution over possible moves) and the value network (gives probability of winning given the board). Blog post here: deepmind.com/blog/alphago-zero-learning-scratch and the paper is linked within. $\endgroup$
    – Kevin Wang
    Mar 26, 2018 at 19:37

7 Answers 7


In brief each net has a different purpose as you mentioned:

  • The value network was used at the leaf nodes to reduce the depth of the tree search.
  • The policy network was used to reduce the breadth of the search from a node (guiding towards promising immediate actions).

In general, you can use value function methods to find an optimal policy or directly search in the policy space to optimize a parametrized policy function (of course there are pros and cons). You can use function approximators (e.g Deep Nets) in each case. I see that mainly you are confused about the policy net so I focus my answer into this.

The policy net was first:

trained to do the moves that most likely a human would do given a board state (so input is a board state and output is a histogram that shows the probability of each action given that state). The net can approximate the probability function underlying the mapping from states to actions. It is reasonable to think to start building your policy from available data after all. After supervised training using experts moves the policy net could play the game sufficient (although far from a Master's level). Simply, you attempted to capture the general pattern of action selection of professional players.


it was trained in games with opponent itself, in order to optimize the previous-learned policy. This time its weights were updated using the REINFORCE algorithm. By doing this, you update the net parameters towards maximization of expected reward. Eventually you have a net that not only selects the actions like a professional player but also towards winning the game (However it cannot plan!).

After this step, they approximated the value function of a bit more noisy version of the learned policy, by regression (input is the state board and target the result of the game). You can use this network to affect the leaf node evaluation.

Conceptually speaking, the policy net gives you a probability over actions, but this doesn't indicate that you will end up in a good, for winning the game, state. AlphaGo had some "blind spots" and during the tournament did some really bad moves but also one exceptional move that a human could never had thought.

Finally you can use your planning algorithm (MCTS) in combination with these nets. Why we took all these steps? Briefly, the simple MCTS without any "intuition" would have failed.

  • $\begingroup$ I was just returning to this question having learnt about Actor-Critic, and how it offsets costs/benefits between pure Value based vs Policy based methods. I think this answer comes closest to explaining this detail, but it is covered very well in David Silver's lecture: www0.cs.ucl.ac.uk/staff/d.silver/web/Teaching_files/pg.pdf (lecture 7, it is also available on YouTube - youtube.com/… ). $\endgroup$ Jul 11, 2016 at 8:32
  • $\begingroup$ Pure MCTS-based approaches have shown to be succesful though, so in a wider context I wouldn't say it would have failed. $\endgroup$
    – Zimano
    Mar 2, 2018 at 18:23
  • 1
    $\begingroup$ I am only referring to the case of the game of Go and also about reaching Master level of game play. Actually David Silver (as other researchers) was working on solving the game of Go with MCTS since his PhD. But nothing could compete at the highest levels till the combination of the methods above. $\endgroup$ Mar 2, 2018 at 20:22

Here is my concise thought process in understanding the two different networks.

First of all, the goal is to find an optimal solution (or very near-optimal) without using an exhaustive search, which is definitely challenging.

Per position or state, there will be N moves possible, and on each move there will be its own depth D in a full search tree. It is theoretically or mathematically possible to walk through all paths and find an optimal solution(s). However, we don't want to do a full search.

Now we got two separate questions for developing an approximation approach.

Q1. How can we skip or disregard some moves out of N per position? (i.e., breath reduction)

Q2. How can we stop at an intermediate depth in a search tree rather than walking through until the end of game, without failing to find an optimal solution? (i.e., depth reduction)

The policy network is mainly designed for filtering out useless moves out of N, yet without failing to find an optimal solution. Here this network initially relies on human expert moves, i.e., SL, and improved by RL later.

The value network is mainly designed for finding the winning probability without a full search.

These two networks have a common goal of finding an optimal solution, However, in each strategic choice of move, each network plays a different role.

I just hope this helps. I know it'd be still at a high level.

  • $\begingroup$ Your two ways explanation is very concise. $\endgroup$
    – Tim
    Jun 2, 2017 at 14:57

I think the OP was confusing about AlphaGo with alpha-beta. In alpha-beta, you'd indeed use the policy network for helping with pruning, but not here. Again, there is no pruning as the algorithm relies on Monte-Carlo tree search (MCTS).

Anyone who thinks my answer is too long might skip to the summary section, where I state why the two networks are not redundant.

In the following example, I'll do some simplification to make my ideas easier to understand.


Imagine you have a position where there are two legal moves. The first move is a dead-lost for you, however, the second move gives you a winning advantage.

  • First move: forced loss for you
  • Second move: forced win for you

Evaluation network

Let's assume the evaluation network Google gives you is perfect. It can evaluate any leaf position in our example perfectly. We won't change our value network in the example.

To simplify our example, let's assume our value network gives:

  • -1000 for any leaf position which is a loss for you
  • +1000 for any leaf position which is a win for you

Policy network

Let's assume Google gives you two policy networks. The probabilities generated for our position is:

  • Policy 1: 0.9 for move 1 and 0.1 for move 2
  • Policy 2: 0.2 for move 1 and 0.8 for move 2.

Note that our first policy network gives incorrect prior probability for our example. It gives 0.9 for move 1, which is a losing move. This is fine because not even Google could train a perfect policy network.

Playing with the first policy network

AlphaGo needs to generate a simulation with Monte-Carlo, and it needs to choose move 1 or 2. Now, AlphaGo draws a uniform-distributed random variable, and it'll pick:

  • Move 1 if the random number is <= 0.9
  • Move 2 if the random number is > 0.9

So AlphaGo is much more likely to pick the losing move to simulate (in our very first simulation). In our first simulation, we will also use the value network to get a score for the simulation. In the paper, it's:

enter image description here

This value would be -1000, because this simulation would lead to a loss.

Now, AlphaGo needs to generate the second simulation. Again, the first move would be much more likely to pick. But eventually, the second move would be pick because:

  • Our prior probability for the second move is 0.1, not zero
  • AlphaGo is encouraged to try moves which haven't been explored much. In the paper, this is done by this equation:

enter image description here

Note that N is the number of moves searched for the move and it's in the denominator. The more likely our first move is searched, the smaller the u function is. Thus, the probability for selecting our second move improves because AlphaGo actually picks a move by this equation:

enter image description here

enter image description here

This is the key equation. Please look at it carefully:

  • It has a term P for the prior probability (given by the policy network)
  • It has a term Q for the evaluation scores (given by the value network)

Now, we know our second move will eventually be chosen. When it does happen, the value network gives a +1000. This will increase Q, which makes the second move much more likely be chosen in the later simulations.

Given enough simulations, the number of times the second move is chosen for simulation should be more than the number of times the first move is chosen.

Finally, the move that AlphaGo decides to make is (quoted from the paper):

Once the search is complete, the algorithm chooses the most visited move from the root position.

Playing with the second policy network

Our second policy network will need less iterations to pick move 2 because it's prior probability given by the policy network is correct in the first place.


Everything here is very similar to Bayesian analysis. We start off with some prior probability (given by the policy network), then we generate data to move the probability distirubtion (given by the value network).


  • Policy network is used to generate prior probabilities to guide what move the Monte-Carlo search should pick
  • Value network is used to generate data to validate the policy network. If the policy network is bad, AlphaGo would need more computing resources to converge (if ever).
  • You can think of it like Bayesian analysis
  • $\begingroup$ I think this gives much deeper insight to the internal mechanisms. I am still not sure whether it explains why the two networks. The issue I have is "assume the evaluation network ... is perfect". If that was the case, then indeed the policy network is redundant. Just look one move ahead (for all possible moves) and pick the one with the best value network assessment. Of course the value network isn't perfect, and I suspect it gets more accurate the further progressed into the game . . . but I don't know how true/useful that is, or whether it completes this answer. $\endgroup$ Apr 2, 2016 at 18:06
  • $\begingroup$ @NeilSlater Ok. The networks aren ot perfect, but the reasons I have here are still good, just that we need more MC simulations. $\endgroup$
    – SmallChess
    Apr 2, 2016 at 22:36
  • $\begingroup$ @StudentT Nice explanation of some of the key equations. I would make one small change: It doesn't make a lot of sense to say "You need to see ten moves ahead" to understand the position in Monte Carlo Tree Search. MCTS is a depth-first proof number search, and we don't really ever reach fixed depths like we would with iterative deepening in chess. Even with the value network allowing us evaluations of nodes before the end of the game, we are still not reaching these in a breadth first manner, and there is no min-max evaluation of the nodes, etc. $\endgroup$
    – Imran
    Apr 4, 2016 at 20:14
  • $\begingroup$ @Imran Feel free to edit my post. $\endgroup$
    – SmallChess
    Apr 5, 2016 at 1:19

Policy Network: The Network which learns to give a definite output by giving a particular Input to the game is known as Policy Network.

Value Networks: The value network assigns value/score to the state of the game by calculating an expected cumulative score for the current state s. Every state goes through the value network. The states which get more reward obviously get more value in the network.

Better understanding with Animations Go here: Policy Networks vs Value Networks in Reinforcement Learning

enter image description here


From what I understand the difference is in the outputs. Where policy network outputs a probability distribution over the possible moves, the value network returns a real value that can be interpreted as the probability of winning given this board configuration. From there Monte-Carlo tree search is performed via taking top K moves from and then narrowing the search tree again by taking top K value network outputs.

Feel obligated to correct me if I am wrong.

  • 2
    $\begingroup$ You're not wrong AFAIK, but I don't think this answers the question. OP seems to be aware of the difference between a value-based or policy-based model in RL. The two approaches do indeed have different outputs. However, it is more common to see one or the other, not both, and the two models are usually redundant in that "Value_maxarg( S' ) == Policy_maxarg( S, A )" where S is current state, A is action to take, and S' is resulting state. I.e. the two networks in a more usual RL setup would simply produce the same result, even though the outputs are different. $\endgroup$ Mar 29, 2016 at 9:04

I just have a very shallow understanding. But I feel in an idea network where you have trained every possible case, one network is enough. But it's not possible to do so in limited resources. So policy to reduce breadth and value to reduce depth. It's a probability result and but not a 100% perfect result.


Yes, Policy-Net & Value-Net are redundant. If you have a Policy-Network you don't need a Value-Network and vice versa. At first glance you might think that these 2 networks would behave exactly the same once fully trained but that ain't the case.

A good example for that is the following problem:

Imagine there are 2 states (S0 and S1) and in S0 the agent can use 2 different actions (a0, a1) which will both lead to S1. In S1 however the Agent is rewarded or punished depending on the action the Agent has taken. If it reached S1 by using a0 it's rewarded and if it reached S1 by using a1 it is punished.

A Value-Network would be helpless in this situation cause it only looks at the states and not the actions that lead to the states. In one pass it would increase the value of S1 and in another pass were it was punished it would decrease the value of S1 again.

A Policy-Network on the other hand just looks at the actions that where taken and ignores the states alltogether. It would pick up the fact that a0 leads to reward and a1 leads to punishment very quickly.

Imho Value-Networks work best with Board-Games and similar problems where each state clearly defines a reward/punishment value. If that's not the case or the number of possible states is too big then a Policy-Network is the way-to-go.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.