I am currently trying to understand how TD-Gammon works and have two questions:
1) I found an article which explains the weight update. It consists of three part. The last part is an differentiation of V(s) with respect to w. In the text it is called a "running sum". How do I calculate that value? (I'm only interested in the weight changes from the output to the hidden layer, not in further weight changes)
2) After having read this procedure of updating the weights, there has one question arised: Why don't we just create a target value for a state using reinforcement learning and give that value to our neural network, so that it learns to return that value for the current state? Why is there an extra updating rule directly manipulating the weights?
Really, you just need to implement an ANN which uses the basic, usual sum of squares error. Then, replace the target network outputs with the TD-error value: E = r + gamma*V(t+1) - V(t)
From there, you can just use the typical ANN backprop weight update rule.
So, in short, I think your description is actually what a RL via ANN algorithm should do. It is training the ANN to learn the state/action value function.
Related
While trying to implement the Episodic Semi-gradient Sarsa with a Neural Network as the approximator I wondered how I choose the optimal action based on the currently learned weights of the network. If the action space is discrete I can just calculate the estimated value of the different actions in the current state and choose the one which gives the maximimum. But this seems to be not the best way of solving the problem. Furthermore, it does not work if the action space can be continous (like the acceleration of a self-driving car for example).
So, basicly I am wondering how to solve the 10th line Choose A' as a function of q(S', , w) in this pseudo-code of Sutton:
How are these problems typically solved? Can one recommend a good example of this algorithm using Keras?
Edit: Do I need to modify the pseudo-code when using a network as the approximator? So, that I simply minimize the MSE of the prediction of the network and the reward R for example?
I wondered how I choose the optimal action based on the currently learned weights of the network
You have three basic choices:
Run the network multiple times, once for each possible value of A' to go with the S' value that you are considering. Take the maximum value as the predicted optimum action (with probability of 1-ε, otherwise choose randomly for ε-greedy policy typically used in SARSA)
Design the network to estimate all action values at once - i.e. to have |A(s)| outputs (perhaps padded to cover "impossible" actions that you need to filter out). This will alter the gradient calculations slightly, there should be zero gradient applied to last layer inactive outputs (i.e. anything not matching the A of (S,A)). Again, just take the maximum valid output as the estimated optimum action. This can be more efficient than running the network multiple times. This is also the approach used by the recent DQN Atari games playing bot, and AlphaGo's policy networks.
Use a policy-gradient method, which works by using samples to estimate gradient that would improve a policy estimator. You can see chapter 13 of Sutton and Barto's second edition of Reinforcement Learning: An Introduction for more details. Policy-gradient methods become attractive for when there are large numbers of possible actions and can cope with continuous action spaces (by making estimates of the distribution function for optimal policy - e.g. choosing mean and standard deviation of a normal distribution, which you can sample from to take your action). You can also combine policy-gradient with a state-value approach in actor-critic methods, which can be more efficient learners than pure policy-gradient approaches.
Note that if your action space is continuous, you don't have to use a policy-gradient method, you could just quantise the action. Also, in some cases, even when actions are in theory continuous, you may find the optimal policy involves only using extreme values (the classic mountain car example falls into this category, the only useful actions are maximum acceleration and maximum backwards acceleration)
Do I need to modify the pseudo-code when using a network as the approximator? So, that I simply minimize the MSE of the prediction of the network and the reward R for example?
No. There is no separate loss function in the pseudocode, such as the MSE you would see used in supervised learning. The error term (often called the TD error) is given by the part in square brackets, and achieves a similar effect. Literally the term ∇q(S,A,w) (sorry for missing hat, no LaTex on SO) means the gradient of the estimator itself - not the gradient of any loss function.
I'm trying to get a better understanding of neural networks by trying to programm a Convolution Neural Network by myself.
So far, I'm going to make it pretty simple by not using max-pooling and using simple ReLu-activation. I'm aware of the disadvantages of this setup, but the point is not making the best image detector in the world.
Now, I'm stuck understanding the details of the error calculation, propagating it back and how it interplays with the used activation-function for calculating the new weights.
I read this document (A Beginner's Guide To Understand CNN), but it doesn't help me understand much. The formula for calculating the error already confuses me.
This sum-function doesn't have defined start- and ending points, so i basically can't read it. Maybe you can simply provide me with the correct one?
After that, the author assumes a variable L that is just "that value" (i assume he means E_total?) and gives an example for how to define the new weight:
where W is the weights of a particular layer.
This confuses me, as i always stood under the impression the activation-function (ReLu in my case) played a role in how to calculate the new weight. Also, this seems to imply i simply use the error for all layers. Doesn't the error value i propagate back into the next layer somehow depends on what i calculated in the previous one?
Maybe all of this is just uncomplete and you can point me into the direction that helps me best for my case.
Thanks in advance.
You do not backpropagate errors, but gradients. The activation function plays a role in caculating the new weight, depending on whether or not the weight in question is before or after said activation, and whether or not it is connected. If a weight w is after your non-linearity layer f, then the gradient dL/dw wont depend on f. But if w is before f, then, if they are connected, then dL/dw will depend on f. For example, suppose w is the weight vector of a fully connected layer, and assume that f directly follows this layer. Then,
dL/dw=(dL/df)*df/dw //notations might change according to the shape
//of the tensors/matrices/vectors you chose, but
//this is just the chain rule
As for your cost function, it is correct. Many people write these formulas in this non-formal style so that you get the idea, but that you can adapt it to your own tensor shapes. By the way, this sort of MSE function is better suited to continous label spaces. You might want to use softmax or an svm loss for image classification (I'll come back to that). Anyway, as you requested a correct form for this function, here is an example. Imagine you have a neural network that predicts a vector field of some kind (like surface normals). Assume that it takes a 2d pixel x_i and predicts a 3d vector v_i for that pixel. Now, in your training data, x_i will already have a ground truth 3d vector (i.e label), that we'll call y_i. Then, your cost function will be (the index i runs on all data samples):
sum_i{(y_i-v_i)^t (y_i-vi)}=sum_i{||y_i-v_i||^2}
But as I said, this cost function works if the labels form a continuous space (here , R^3). This is also called a regression problem.
Here's an example if you are interested in (image) classification. I'll explain it with a softmax loss, the intuition for other losses is more or less similar. Assume we have n classes, and imagine that in your training set, for each data point x_i, you have a label c_i that indicates the correct class. Now, your neural network should produce scores for each possible label, that we'll note s_1,..,s_n. Let's note the score of the correct class of a training sample x_i as s_{c_i}. Now, if we use a softmax function, the intuition is to transform the scores into a probability distribution, and maximise the probability of the correct classes. That is , we maximse the function
sum_i { exp(s_{c_i}) / sum_j(exp(s_j))}
where i runs over all training samples, and j=1,..n on all class labels.
Finally, I don't think the guide you are reading is a good starting point. I recommend this excellent course instead (essentially the Andrew Karpathy parts at least).
I'm trying to navigate an agent in a n*n gridworld domain by using Q-Learning + a feedforward neural network as a q-function approximator. Basically the agent should find the best/shortest way to reach a certain terminal goal position (+10 reward). Every step the agent takes it gets -1 reward. In the gridworld there are also some positions the agent should avoid (-10 reward, terminal states,too).
So far I implemented a Q-learning algorithm, that saves all Q-values in a Q-table and the agent performs well.
In the next step, I want to replace the Q-table by a neural network, trained online after every step of the agent. I tried a feedforward NN with one hidden layer and four outputs, representing the Q-values for the possible actions in the gridworld (north,south,east, west).
As input I used a nxn zero-matrix, that has a "1" at the current positions of the agent.
To reach my goal I tried to solve the problem from the ground up:
Explore the gridworld with standard Q-Learning and use the Q-map as training data for the Network once Q-Learning is finished
--> worked fine
Use Q-Learning and provide the updates of the Q-map as trainingdata
for NN (batchSize = 1)
--> worked good
Replacy the Q-Map completely by the NN. (This is the point, when it gets interesting!)
-> FIRST MAP: 4 x 4
As described above, I have 16 "discrete" Inputs, 4 Output and it works fine with 8 neurons(relu) in the hidden layer (learning rate: 0.05). I used a greedy policy with an epsilon, that reduces from 1 to 0.1 within 60 episodes.
The test scenario is shown here. Performance is compared beetween standard qlearning with q-map and "neural" qlearning (in this case i used 8 neurons and differnt dropOut rates).
To sum it up: Neural Q-learning works good for small grids, also the performance is okay and reliable.
-> Bigger MAP: 10 x 10
Now I tried to use the neural network for bigger maps.
At first I tried this simple case.
In my case the neural net looks as following: 100 input; 4 Outputs; about 30 neurons(relu) in one hidden layer; again I used a decreasing exploring factor for greedy policy; over 200 episodes the learning rate decreases from 0.1 to 0.015 to increase stability.
At frist I had problems with convergence and interpolation between single positions caused by the discrete input vector.
To solve this I added some neighbour positions to the vector with values depending on thier distance to the current position. This improved the learning a lot and the policy got better. Performance with 24 neurons is seen in the picture above.
Summary: the simple case is solved by the network, but only with a lot of parameter tuning (number of neurons, exploration factor, learning rate) and special input transformation.
Now here are my questions/problems I still haven't solved:
(1) My network is able to solve really simple cases and examples in a 10 x 10 map, but it fails as the problem gets a bit more complex. In cases where failing is very likely, the network has no change to find a correct policy.
I'm open minded for any idea that could improve performace in this cases.
(2) Is there a smarter way to transform the input vector for the network? I'm sure that adding the neighboring positons to the input vector on the one hand improve the interpolation of the q-values over the map, but on the other hand makes it harder to train special/important postions to the network. I already tried standard cartesian two-dimensional input (x/y) on an early stage, but failed.
(3) Is there another network type than feedforward network with backpropagation, that generally produces better results with q-function approximation? Have you seen projects, where a FF-nn performs well with bigger maps?
It's known that Q-Learning + a feedforward neural network as a q-function approximator can fail even in simple problems [Boyan & Moore, 1995].
Rich Sutton has a question in the FAQ of his web site related with this.
A possible explanation is the phenomenok known as interference described in [Barreto & Anderson, 2008]:
Interference happens when the update of one state–action pair changes the Q-values of other pairs, possibly in the wrong direction.
Interference is naturally associated with generalization, and also happens in conventional supervised learning. Nevertheless, in the reinforcement learning paradigm its effects tend to be much more harmful. The reason for this is twofold. First, the combination of interference and bootstrapping can easily become unstable, since the updates are no longer strictly local. The convergence proofs for the algorithms derived from (4) and (5) are based on the fact that these operators are contraction mappings, that is, their successive application results in a sequence converging to a fixed point which is the solution for the Bellman equation [14,36]. When using approximators, however, this asymptotic convergence is lost, [...]
Another source of instability is a consequence of the fact that in on-line reinforcement learning the distribution of the incoming data depends on the current policy. Depending on the dynamics of the system, the agent can remain for some time in a region of the state space which is not representative of the entire domain. In this situation, the learning algorithm may allocate excessive resources of the function approximator to represent that region, possibly “forgetting” the previous stored information.
One way to alleviate the interference problem is to use a local function approximator. The more independent each basis function is from each other, the less severe this problem is (in the limit, one has one basis function for each state, which corresponds to the lookup-table case) [86]. A class of local functions that have been widely used for approximation is the radial basis functions (RBFs) [52].
So, in your kind of problem (n*n gridworld), an RBF neural network should produce better results.
References
Boyan, J. A. & Moore, A. W. (1995) Generalization in reinforcement learning: Safely approximating the value function. NIPS-7. San Mateo, CA: Morgan Kaufmann.
André da Motta Salles Barreto & Charles W. Anderson (2008) Restricted gradient-descent algorithm for value-function approximation in reinforcement learning, Artificial Intelligence 172 (2008) 454–482
I am trying to extract common patterns that always appear whenever a certain event occurs.
For example, patient A, B, and C all had a heart attack. Using the readings from there pulse, I want to find the common patterns before the heart attack stroke.
In the next stage I want to do this using multiple dimensions. For example, using the readings from the patients pulse, temperature, and blood pressure, what are the common patterns that occurred in the three dimensions taking into consideration the time and order between each dimension.
What is the best way to solve this problem using Neural Networks and which type of network is best?
(Just need some pointing in the right direction)
and thank you all for reading
Described problem looks like a time series prediction problem. That means a basic prediction problem for a continuous or discrete phenomena generated by some existing process. As a raw data for this problem we will have a sequence of samples x(t), x(t+1), x(t+2), ..., where x() means an output of considered process and t means some arbitrary timepoint.
For artificial neural networks solution we will consider a time series prediction, where we will organize our raw data to a new sequences. As you should know, we consider X as a matrix of input vectors that will be used in ANN learning. For time series prediction we will construct a new collection on following schema.
In the most basic form your input vector x will be a sequence of samples (x(t-k), x(t-k+1), ..., x(t-1), x(t)) taken at some arbitrary timepoint t, appended to it predecessor samples from timepoints t-k, t-k+1, ..., t-1. You should generate every example for every possible timepoint t like this.
But the key is to preprocess data so that we get the best prediction results.
Assuming your data (phenomena) is continuous, you should consider to apply some sampling technique. You could start with an experiment for some naive sampling period Δt, but there are stronger methods. See for example Nyquist–Shannon Sampling Theorem, where the key idea is to allow to recover continuous x(t) from discrete x(Δt) samples. This is reasonable when we consider that we probably expect our ANNs to do this.
Assuming your data is discrete... you still should need to try sampling, as this will speed up your computations and might possibly provide better generalization. But the key advice is: do experiments! as the best architecture depends on data and also will require to preprocess them correctly.
The next thing is network output layer. From your question, it appears that this will be a binary class prediction. But maybe a wider prediction vector is worth considering? How about to predict the future of considered samples, that is x(t+1), x(t+2) and experiment with different horizons (length of the future)?
Further reading:
Somebody mentioned Python here. Here is some good tutorial on timeseries prediction with Keras: Victor Schmidt, Keras recurrent tutorial, Deep Learning Tutorials
This paper is good if you need some real example: Fessant, Francoise, Samy Bengio, and Daniel Collobert. "On the prediction of solar activity using different neural network models." Annales Geophysicae. Vol. 14. No. 1. 1996.
I'm trying to create a sample neural network that can be used for credit scoring. Since this is a complicated structure for me, i'm trying to learn them small first.
I created a network using back propagation - input layer (2 nodes), 1 hidden layer (2 nodes +1 bias), output layer (1 node), which makes use of sigmoid as activation function for all layers. I'm trying to test it first using a^2+b2^2=c^2 which means my input would be a and b, and the target output would be c.
My problem is that my input and target output values are real numbers which can range from (-/infty, +/infty). So when I'm passing these values to my network, my error function would be something like (target- network output). Would that be correct or accurate? In the sense that I'm getting the difference between the network output (which is ranged from 0 to 1) and the target output (which is a large number).
I've read that the solution would be to normalise first, but I'm not really sure how to do this. Should i normalise both the input and target output values before feeding them to the network? What normalisation function is best to use cause I read different methods in normalising. After getting the optimized weights and use them to test some data, Im getting an output value between 0 and 1 because of the sigmoid function. Should i revert the computed values to the un-normalized/original form/value? Or should i only normalise the target output and not the input values? This really got me stuck for weeks as I'm not getting the desired outcome and not sure how to incorporate the normalisation idea in my training algorithm and testing..
Thank you very much!!
So to answer your questions :
Sigmoid function is squashing its input to interval (0, 1). It's usually useful in classification task because you can interpret its output as a probability of a certain class. Your network performes regression task (you need to approximate real valued function) - so it's better to set a linear function as an activation from your last hidden layer (in your case also first :) ).
I would advise you not to use sigmoid function as an activation function in your hidden layers. It's much better to use tanh or relu nolinearities. The detailed explaination (as well as some useful tips if you want to keep sigmoid as your activation) might be found here.
It's also important to understand that architecture of your network is not suitable for a task which you are trying to solve. You can learn a little bit of what different networks might learn here.
In case of normalization : the main reason why you should normalize your data is to not giving any spourius prior knowledge to your network. Consider two variables : age and income. First one varies from e.g. 5 to 90. Second one varies from e.g. 1000 to 100000. The mean absolute value is much bigger for income than for age so due to linear tranformations in your model - ANN is treating income as more important at the beginning of your training (because of random initialization). Now consider that you are trying to solve a task where you need to classify if a person given has grey hair :) Is income truly more important variable for this task?
There are a lot of rules of thumb on how you should normalize your input data. One is to squash all inputs to [0, 1] interval. Another is to make every variable to have mean = 0 and sd = 1. I usually use second method when the distribiution of a given variable is similiar to Normal Distribiution and first - in other cases.
When it comes to normalize the output it's usually also useful to normalize it when you are solving regression task (especially in multiple regression case) but it's not so crucial as in input case.
You should remember to keep parameters needed to restore the original size of your inputs and outputs. You should also remember to compute them only on a training set and apply it on both training, test and validation sets.