马尔可夫决策过程：价值迭代，它是如何运作的？

Question

I've been reading a lot about Markov Decision Processes (using value iteration) lately but I simply can't get my head around them. I've found a lot of resources on the Internet / books, but they all use mathematical formulas that are way too complex for my competencies.

Since this is my first year at college, I've found that the explanations and formulas provided on the web use notions / terms that are way too complicated for me and they assume that the reader knows certain things that I've simply never heard of.

I want to use it on a 2D grid (filled with walls(unattainable), coins(desirable) and enemies that move(which must be avoided at all costs)). The whole goal is to collect all the coins without touching the enemies, and I want to create an AI for the main player using a Markov Decision Process ( MDP ). Here is how it partially looks like (note that the game-related aspect is not so much of a concern here. I just really want to understand MDPs in general):

From what I understand, a rude simplification of MDPs is that they can create a grid which holds in which direction we need to go (kind of a grid of "arrows" pointing where we need to go, starting at a certain position on the grid) to get to certain goals and avoid certain obstacles. Specific to my situation, that would mean that it allows the player to know in which direction to go to collect the coins and avoid the enemies.

Now, using the MDP terms, it would mean that it creates a collection of states(the grid) which holds certain policies(the action to take -> up, down, right, left) for a certain state(a position on the grid). The policies are determined by the "utility" values of each state, which themselves are calculated by evaluating how much getting there would be beneficial in the short and long term.

Is this correct? Or am I completely on the wrong track?

I'd at least like to know what the variables from the following equation represent in my situation:

(taken from the book "Artificial Intelligence - A Modern Approach" from Russell & Norvig)

I know that s would be a list of all the squares from the grid, a would be a specific action (up / down / right / left), but what about the rest?

How would the reward and utility functions be implemented?

It would be really great if someone knew a simple link which shows pseudo-code to implement a basic version with similarities to my situation in a very slow way, because I don't even know where to start here.

Thank you for your precious time.

(Note: feel free to add / remove tags or tell me in the comments if I should give more details about something or anything like that.)

Answer 1

Yes, the mathematical notation can make it seem much more complicated than it is. Really, it is a very simple idea. I have a implemented a value iteration demo applet that you can play with to get a better idea.

Basically, lets say you have a 2D grid with a robot in it. The robot can try to move North, South, East, West (those are the actions a) but, because its left wheel is slippery, when it tries to move North there is only a .9 probability that it will end up at the square North of it while there is a .1 probability that it will end up at the square West of it (similarly for the other 3 actions). These probabilities are captured by the T() function. Namely, T(s,A,s') will look like:

s    A      s'     T    //x=0,y=0 is at the top-left of the screen
x,y  North  x,y+1  .9   //we do move north
x,y  North  x-1,y  .1   //wheels slipped, so we move West
x,y  East   x+1,y  .9
x,y  East   x,y-1  .1
x,y  South  x,y+1  .9
x,y  South  x-1,y  .1 
x,y  West   x-1,y  .9
x,y  West   x,y+1  .1 

You then set the Reward to be 0 for all states, but 100 for the goal state, that is, the location you want the robot to get to.

What value-iteration does is its starts by giving a Utility of 100 to the goal state and 0 to all the other states. Then on the first iteration this 100 of utility gets distributed back 1-step from the goal, so all states that can get to the goal state in 1 step (all 4 squares right next to it) will get some utility. Namely, they will get a Utility equal to the probability that from that state we can get to the goal stated. We then continue iterating, at each step we move the utility back 1 more step away from the goal.

In the example above, say you start with R(5,5)= 100 and R(.) = 0 for all other states. So the goal is to get to 5,5.

On the first iteration we set

R(5,6) = gamma * (.9 * 100) + gamma * (.1 * 100)

because on 5,6 if you go North there is a .9 probability of ending up at 5,5, while if you go West there is a .1 probability of ending up at 5,5.

Similarly for (5,4), (4,5), (6,5).

All other states remain with U = 0 after the first iteration of value iteration.

Answer 2

I would recommend using Q-learning for your implementation.

Maybe you can use this post I wrote as an inspiration. This is a Q-learning demo with Java source code . This demo is a map with 6 fields and the AI learns where it should go from every state to get to the reward.

Q-learning is a technique for letting the AI learn by itself by giving it reward or punishment.

This example shows the Q-learning used for path finding. A robot learns where it should go from any state.

The robot starts at a random place, it keeps memory of the score while it explores the area, whenever it reaches the goal, we repeat with a new random start. After enough repetitions the score values will be stationary (convergence).

In this example the action outcome is deterministic (transition probability is 1) and the action selection is random. The score values are calculated by the Q-learning algorithm Q(s,a).
The image shows the states (A,B,C,D,E,F), possible actions from the states and the reward given.

Result Q*(s,a)

Policy Π*(s)

Qlearning.java

 import java.text.DecimalFormat; import java.util.Random; /** * @author Kunuk Nykjaer */ public class Qlearning { final DecimalFormat df = new DecimalFormat("#.##"); // path finding final double alpha = 0.1; final double gamma = 0.9; // states A,B,C,D,E,F // eg from A we can go to B or D // from C we can only go to C // C is goal state, reward 100 when B->C or F->C // // _______ // |A|B|C| // |_____| // |D|E|F| // |_____| // final int stateA = 0; final int stateB = 1; final int stateC = 2; final int stateD = 3; final int stateE = 4; final int stateF = 5; final int statesCount = 6; final int[] states = new int[]{stateA,stateB,stateC,stateD,stateE,stateF}; // http://en.wikipedia.org/wiki/Q-learning // http://people.revoledu.com/kardi/tutorial/ReinforcementLearning/Q-Learning.htm // Q(s,a)= Q(s,a) + alpha * (R(s,a) + gamma * Max(next state, all actions) - Q(s,a)) int[][] R = new int[statesCount][statesCount]; // reward lookup double[][] Q = new double[statesCount][statesCount]; // Q learning int[] actionsFromA = new int[] { stateB, stateD }; int[] actionsFromB = new int[] { stateA, stateC, stateE }; int[] actionsFromC = new int[] { stateC }; int[] actionsFromD = new int[] { stateA, stateE }; int[] actionsFromE = new int[] { stateB, stateD, stateF }; int[] actionsFromF = new int[] { stateC, stateE }; int[][] actions = new int[][] { actionsFromA, actionsFromB, actionsFromC, actionsFromD, actionsFromE, actionsFromF }; String[] stateNames = new String[] { "A", "B", "C", "D", "E", "F" }; public Qlearning() { init(); } public void init() { R[stateB][stateC] = 100; // from b to c R[stateF][stateC] = 100; // from f to c } public static void main(String[] args) { long BEGIN = System.currentTimeMillis(); Qlearning obj = new Qlearning(); obj.run(); obj.printResult(); obj.showPolicy(); long END = System.currentTimeMillis(); System.out.println("Time: " + (END - BEGIN) / 1000.0 + " sec."); } void run() { /* 1. Set parameter , and environment reward matrix R 2. Initialize matrix Q as zero matrix 3. For each episode: Select random initial state Do while not reach goal state o Select one among all possible actions for the current state o Using this possible action, consider to go to the next state o Get maximum Q value of this next state based on all possible actions o Compute o Set the next state as the current state */ // For each episode Random rand = new Random(); for (int i = 0; i < 1000; i++) { // train episodes // Select random initial state int state = rand.nextInt(statesCount); while (state != stateC) // goal state { // Select one among all possible actions for the current state int[] actionsFromState = actions[state]; // Selection strategy is random in this example int index = rand.nextInt(actionsFromState.length); int action = actionsFromState[index]; // Action outcome is set to deterministic in this example // Transition probability is 1 int nextState = action; // data structure // Using this possible action, consider to go to the next state double q = Q(state, action); double maxQ = maxQ(nextState); int r = R(state, action); double value = q + alpha * (r + gamma * maxQ - q); setQ(state, action, value); // Set the next state as the current state state = nextState; } } } double maxQ(int s) { int[] actionsFromState = actions[s]; double maxValue = Double.MIN_VALUE; for (int i = 0; i < actionsFromState.length; i++) { int nextState = actionsFromState[i]; double value = Q[s][nextState]; if (value > maxValue) maxValue = value; } return maxValue; } // get policy from state int policy(int state) { int[] actionsFromState = actions[state]; double maxValue = Double.MIN_VALUE; int policyGotoState = state; // default goto self if not found for (int i = 0; i < actionsFromState.length; i++) { int nextState = actionsFromState[i]; double value = Q[state][nextState]; if (value > maxValue) { maxValue = value; policyGotoState = nextState; } } return policyGotoState; } double Q(int s, int a) { return Q[s][a]; } void setQ(int s, int a, double value) { Q[s][a] = value; } int R(int s, int a) { return R[s][a]; } void printResult() { System.out.println("Print result"); for (int i = 0; i < Q.length; i++) { System.out.print("out from " + stateNames[i] + ": "); for (int j = 0; j < Q[i].length; j++) { System.out.print(df.format(Q[i][j]) + " "); } System.out.println(); } } // policy is maxQ(states) void showPolicy() { System.out.println("\\nshowPolicy"); for (int i = 0; i < states.length; i++) { int from = states[i]; int to = policy(from); System.out.println("from "+stateNames[from]+" goto "+stateNames[to]); } } } 

Print result

 out from A: 0 90 0 72,9 0 0 out from B: 81 0 100 0 81 0 out from C: 0 0 0 0 0 0 out from D: 81 0 0 0 81 0 out from E: 0 90 0 72,9 0 90 out from F: 0 0 100 0 81 0 showPolicy from a goto B from b goto C from c goto C from d goto A from e goto B from f goto C Time: 0.025 sec. 

Answer 3

Not a complete answer, but a clarifying remark.

The state is not a single cell. The state contains the information what is in each cell for all concerned cells at once. This means one state element contains the information which cells are solid and which are empty; which ones contain monsters; where are coins; where is the player.

Maybe you could use a map from each cell to its content as state. This does ignore the movement of monsters and player, which are probably very important, too.

The details depend on how you want to model your problem (deciding what belongs to the state and in which form).

Then a policy maps each state to an action like left, right, jump, etc.

First you must understand the problem that is expressed by a MDP before thinking about how algorithms like value iteration work.

Answer 4

I know this is a fairly old post, but i came across it when looking for MDP related questions, I did want to note (for folks coming in here) a few more comments about when you stated what "s" and "a" were.

I think for a you are absolutely correct it's your list of [up,down,left,right].

However for s it's really the location in the grid and s' is the location you can go to. What that means is that you pick a state, and then you pick a particular s' and go through all the actions that can take you to that sprime, which you use to figure out those values. (pick a max out of those). Finally you go for the next s' and do the same thing, when you've exhausted all the s' values then you find the max of what you just finished searching on.

Suppose you picked a grid cell in the corner, you'd only have 2 states you could possibly move to (assuming bottom left corner), depending on how you choose to "name" your states, we could in this case assume a state is an x,y coordinate, so your current state s is 1,1 and your s' (or s prime) list is x+1,y and x,y+1 (no diagonal in this example) (The Summation part that goes over all s')

Also you don't have it listed in your equation, but the max is of a or the action that gives you the max, so first you pick the s' that gives you the max and then within that you pick the action (at least this is my understanding of the algorithm).

So if you had

x,y+1 left = 10 
x,y+1 right = 5 

x+1,y left = 3
x+1,y right 2

You'll pick x,y+1 as your s', but then you'll need to pick an action that is maximized which is in this case left for x,y+1. I'm not sure if there is a subtle difference between just finding the maximum number and finding the state then the maximum number though so maybe someone someday can clarify that for me.

If your movements are deterministic (meaning if you say go forward, you go forward with 100% certainty), then it's pretty easy you have one action, However if they are non deterministic, you have a say 80% certainty then you should consider the other actions which could get you there. This is the context of the slippery wheel that Jose mentioned above.

I don't want to detract what others have said, but just to give some additional information.