I have a confusion about the way the LSTM networks work when forecasting with an horizon that is not finite but I'm rather searching for a prediction in whatever time in future. In physical terms I would call it the evolution of the system.
Suppose I have a time series $y(t)$ (output) I want to forecast, and some external inputs $u_1(t), u_2(t),\cdots u_N(t)$ on which the series $y(t)$ depends.
It's common to use the lagged value of the output $y(t)$ as input for the network, such that I schematically have something like (let's consider for simplicity just lag 1 for the output and no lag for the external input):
[y(t-1), u_1(t), u_2(t),\cdots u_N(t)] \to y(t)
In this way of thinking the network, when one wants to do recursive forecast it is forced to use the predicted value at the previous step as input for the next step. In this way we have an effect of propagation of error that makes the long term forecast badly behaving.
Now, my confusion is, I'm thinking as a RNN as a kind of an (simple version) implementation of a state space model where I have the inputs, my output and one or more state variable responsible for the memory of the system. These variables are hidden and not observed.
So now the question, if there is this kind of variable taking already into account previous states of the system why would I need to use the lagged output value as input of my network/model ?
Getting rid of this does my long term forecast would be better, since I'm not expecting anymore the propagation of the error of the forecasted output. (I guess there will be anyway an error in the internal state propagating)
Thanks !
Please see DeepAR - a LSTM forecaster more than one step into the future.
The main contributions of the paper are twofold: (1) we propose an RNN
architecture for probabilistic forecasting, incorporating a negative
Binomial likelihood for count data as well as special treatment for
the case when the magnitudes of the time series vary widely; (2) we
demonstrate empirically on several real-world data sets that this
model produces accurate probabilistic forecasts across a range of
input characteristics, thus showing that modern deep learning-based
approaches can effective address the probabilistic forecasting
problem, which is in contrast to common belief in the field and the
mixed results
In this paper, they forecast multiple steps into the future, to negate exactly what you state here which is the error propagation.
Skipping several steps allows to get more accurate predictions, further into the future.
One more thing done in this paper is predicting percentiles, and interpolating, rather than predicting the value directly. This adds stability, and an error assessment.
Disclaimer - I read an older version of this paper.
Related
Would this be a valid implementation of a cross entropy loss that takes the ordinal structure of the GT y into consideration? y_hat is the prediction from a neural network.
ce_loss = F.cross_entropy(y_hat, y, reduction="none")
distance_weight = torch.abs(y_hat.argmax(1) - y) + 1
ordinal_ce_loss = torch.mean(distance_weight * ce_loss)
I'll attempt to answer this question by first fully defining the task, since the question is a bit sparse on details.
I have a set of ordinal classes (e.g. first, second, third, fourth,
etc.) and I would like to predict the class of each data example from
among this set. I would like to define an entropy-based loss-function
for this problem. I would like this loss function to weight the loss
between a predicted class torch.argmax(y_hat) and the true class y
according to the ordinal distance between the two classes. Does the
given loss expression accomplish this?
Short answer: sure, it is "valid". You've roughly implemented L1-norm ordinal class weighting. I'd question whether this is truly the correct weighting strategy for this problem.
For instance, consider that for a true label n, the bin n response is weighted by 1, but the bin n+1 and n-1 responses are weighted by 2. This means that a lot more emphasis will be placed on NOT predicting false positives than on correctly predicting true positives, which may imbue your model with some strange bias.
It also means that examples on the edge will result in a larger total sum of weights, meaning that you'll be weighting examples where the true label is say "first" or "last" more highly than the intermediate classes. (Say you have 5 classes: 1,2,3,4,5. A true label of 1 will require distance_weight of [1,2,3,4,5], the sum of which is 15. A true label of 3 will require distance_weight of [3,2,1,2,3], the sum of which is 11.
In general, classification problems and entropy-based losses are underpinned by the assumption that no set of classes or categories is any more or less related than any other set of classes. In essence, the input data is embedded into an orthogonal feature space where each class represents one vector in the basis. This is quite plainly a bad assumption in your case, meaning that this embedding space is probably not particularly elegant: thus, you have to correct for it with sort of a hack-y weight fix. And in general, this assumption of class non-correlation is probably not true in a great many classification problems (consider e.g. the classic ImageNet classification problem, wherein the class pairs [bus,car], and [bus,zebra] are treated as equally dissimilar. But this is probably a digression into the inherent lack of usefulness of strict ontological structuring of information which is outside the scope of this answer...)
Long Answer: I'd highly suggest moving into a space where the ordinal value you care about is instead expressed in a continuous space. (In the first, second, third example, you might for instance output a continuous value over the range [1,max_place]. This allows you to benefit from loss functions that already capture well the notion that predictions closer in an ordered space are better than predictions farther away in an ordered space (e.g. MSE, Smooth-L1, etc.)
Let's consider one more time the case of the [first,second,third,etc.] ordinal class example, and say that we are trying to predict the places of a set of runners in a race. Consider two races, one in which the first place runner wins by 30% relative to the second place runner, and the second in which the first place runner wins by only 1%. This nuance is entirely discarded by the ordinal discrete classification. In essence, the selection of an ordinal set of classes truncates the amount of information conveyed in the prediction, which means not only that the final prediction is less useful, but also that the loss function encodes this strange truncation and binarization, which is then reflected (perhaps harmfully) in the learned model. This problem could likely be much more elegantly solved by regressing the finishing position, or perhaps instead by regressing the finishing time, of each athlete, and then performing the final ordinal classification into places OUTSIDE of the network training.
In conclusion, you might expect a well-trained ordinal classifier to produce essentially a normal distribution of responses across the class bins, with the distribution peak on the true value: a binned discretization of a space that almost certainly could, and likely should, be treated as a continuous space.
I am currently designing a artificial neural network for a problem with a decay curve.
For example, building a model for predicting the durability of the some material. It may includes the environment condition like temperature and humidity.
However, it is not adequate to predict the durability of the material. For such a problem, I think it is better to using the output durability of previous time slots as one of the current input to predict the durability of next time slot.
Moreover, I do not know how to train a model which feed the output back to input as one of the input columns has only the initial value before training.
For this case,
Method 1 (fail)
I have tried to fill the predicted output durability of current row to the input durability of next row. Nevertheless, it will prevent the model from "loss.backward()" so we cannot compute and update the gradient if we do so. The gradient function used was "CopySlices" instead of "MSELoss" when I copied the predicted output to the next row of the input data.
Feed output to input
gradient function -copy-
Method 2 "fill the input column with expected output"
In this method, I fill the blank input column with expected output (row-1) before training the model. Filling the input column with expected output of previous row is only done for training. For real prediction, I will feed the predicted output to the input. In this case, I am successful to train a overfitting model with MSELoss.
Moreover, I do not believe it is a right method as it uses the expected output as the input no matter how bad it predict. I strongly believed that it is not a right method.
Therefore, I want to ask whether it is possible to feed output to input in linear regression problem using artificial neural network.
I apologize for uploading no code here as I am not convenient to upload the full code here. It may be confidential.
It looks like you need an RNN (recurrent neural network). This tutorial is pretty helpful for understanding an RNN: https://colah.github.io/posts/2015-08-Understanding-LSTMs/.
Learner might be in training stage, where it update Q-table for bunch of epoch.
In this stage, Q-table would be updated with gamma(discount rate), learning rate(alpha), and action would be chosen by random action rate.
After some epoch, when reward is getting stable, let me call this "training is done". Then do I have to ignore these parameters(gamma, learning rate, etc) after that?
I mean, in training stage, I got an action from Q-table like this:
if rand_float < rar:
action = rand.randint(0, num_actions - 1)
else:
action = np.argmax(Q[s_prime_as_index])
But after training stage, Do I have to remove rar, which means I have to get an action from Q-table like this?
action = np.argmax(self.Q[s_prime])
Once the value function has converged (values stop changing), you no longer need to run Q-value updates. This means gamma and alpha are no longer relevant, because they only effect updates.
The epsilon parameter is part of the exploration policy (e-greedy) and helps ensure that the agent visits all states infinitely many times in the limit. This is an important factor in ensuring that the agent's value function eventually converges to the correct value. Once we've deemed the value function converged however, there's no need to continue randomly taking actions that our value function doesn't believe to be best; we believe that the value function is optimal, so we extract the optimal policy by greedily choosing what it says is the best action in every state. We can just set epsilon to 0.
Although the answer provided by #Nick Walker is correct, here it's some additional information.
What you are talking about is closely related with the concept technically known as "exploration-exploitation trade-off". From Sutton & Barto book:
The agent has to exploit what it already knows in order to obtain
reward, but it also has to explore in order to make better action
selections in the future. The dilemma is that neither exploration nor
exploitation can be pursued exclusively without failing at the task.
The agent must try a variety of actions and progressively favor those
that appear to be best.
One way to implement the exploration-exploitation trade-off is using epsilon-greedy exploration, that is what you are using in your code sample. So, at the end, once the agent has converged to the optimal policy, the agent must select only those that exploite the current knowledge, i.e., you can forget the rand_float < rar part. Ideally you should decrease the epsilon parameters (rar in your case) with the number of episodes (or steps).
On the other hand, regarding the learning rate, it worths noting that theoretically this parameter should follow the Robbins-Monro conditions:
This means that the learning rate should decrease asymptotically. So, again, once the algorithm has converged you can (or better, you should) safely ignore the learning rate parameter.
In practice, sometimes you can simply maintain a fixed epsilon and alpha parameters until your algorithm converges and then put them as 0 (i.e., ignore them).
I am just learning Kalman filter. In the Kalman Filter terminology, I am having some difficulty with process noise. Process noise seems to be ignored in many concrete examples (most focused on measurement noise). If someone can point me to some introductory level link that described process noise well with examples, that’d be great.
Let’s use a concrete scalar example for my question, given:
x_j = a x_j-1 + b u_j + w_j
Let’s say x_j models the temperature within a fridge with time. It is 5 degrees and should stay that way, so we model with a = 1. If at some point t = 100, the temperature of the fridge becomes 7 degrees (ie. hot day, poor insulation), then I believe the process noise at this point is 2 degrees. So our state variable x_100 = 7 degrees, and this is the true value of the system.
Question 1:
If I then paraphrase the phrase I often see for describing Kalman filter, “we filter the signal x so that the effects of the noise w are minimized “, http://www.swarthmore.edu/NatSci/echeeve1/Ref/Kalman/ScalarKalman.html if we minimize the effects of the 2 degrees, are we trying to get rid of the 2 degree difference? But the true state at is x_100 == 7 degrees. What are we doing to the process noise w exactly when we Kalmen filter?
Question 2:
The process noise has a variance of Q. In the simple fridge example, it seems easy to model because you know the underlying true state is 5 degrees and you can take Q as the deviation from that state. But if the true underlying state is fluctuating with time, when you model, what part of this would be considered state fluctuation vs. “process noise”. And how do we go about determining a good Q (again example would be nice)?
I have found that as Q is always added to the covariance prediction no matter which time step you are at, (see Covariance prediction formula from http://greg.czerniak.info/guides/kalman1/) that if you select an overly large Q, then it doesn’t seem like the Kalman filter would be well-behaved.
Thanks.
EDIT1 My Interpretation
My interpretation of the term process noise is the difference between the actual state of the system and the state modeled from the state transition matrix (ie. a * x_j-1). And what Kalman filter tries to do, is to bring the prediction closer to the actual state. In that sense, it actually partially "incorporate" the process noise into the prediction through the residual feedback mechanism, rather than "eliminate" it, so that it can predict the actual state better. I have not read such an explanation anywhere in my search, and I would appreciate anyone commenting on this view.
In Kalman filtering the "process noise" represents the idea/feature that the state of the system changes over time, but we do not know the exact details of when/how those changes occur, and thus we need to model them as a random process.
In your refrigerator example:
the state of the system is the temperature,
we obtain measurements of the temperature on some time interval, say hourly,
by looking the thermometer dial. Note that you usually need to
represent the uncertainties involved in the measurement process
in Kalman filtering, but you didn't focus on this in your question.
Let's assume that these errors are small.
At time t you look at the thermometer, see that it says 7degrees;
since we've assumed the measurement errors are very small, that means
that the true temperature is (very close to) 7 degrees.
Now the question is: what is the temperature at some later time, say 15 minutes
after you looked?
If we don't know if/when the condenser in the refridgerator turns on we could have:
1. the temperature at the later time is yet higher than 7degrees (15 minutes manages
to get close to the maximum temperature in a cycle),
2. Lower if the condenser is/has-been running, or even,
3. being just about the same.
This idea that there are a distribution of possible outcomes for the real state of the
system at some later time is the "process noise"
Note: my qualitative model for the refrigerator is: the condenser is not running, the temperature goes up until it reaches a threshold temperature a few degrees above the nominal target temperature (note - this is a sensor so there may be noise in terms of the temperature at which the condenser turns on), the condenser stays on until the temperature
gets a few degrees below the set temperature. Also note that if someone opens the door, then there will be a jump in the temperature; since we don't know when someone might do this, we model it as a random process.
Yeah, I don't think that sentence is a good one. The primary purpose of a Kalman filter is to minimize the effects of observation noise, not process noise. I think the author may be conflating Kalman filtering with Kalman control (where you ARE trying to minimize the effect of process noise).
The state does not "fluctuate" over time, except through the influence of process noise.
Remember, a system does not generally have an inherent "true" state. A refrigerator is a bad example, because it's already a control system, with nonlinear properties. A flying cannonball is a better example. There is some place where it "really is", but that's not intrinsic to A. In this example, you can think of wind as a kind of "process noise". (Not a great example, since it's not white noise, but work with me here.) The wind is a 3-dimensional process noise affecting the cannonball's velocity; it does not directly affect the cannonball's position.
Now, suppose that the wind in this area always blows northwest. We should see a positive covariance between the north and west components of wind. A deviation of the cannonball's velocity northwards should make us expect to see a similar deviation to westward, and vice versa.
Think of Q more as covariance than as variance; the autocorrelation aspect of it is almost incidental.
Its a good discussion going over here. I would like to add that the concept of process noise is that what ever prediction that is made based on the model is having some errors and it is represented using the Q matrix. If you note the equations in KF for prediction of Covariance matrix (P_prediction) which is actually the mean squared error of the state being predicted, the Q is simply added to it. PPredict=APA'+Q . I suggest, it would give a good insight if you could find the derivation of KF equations.
If your state-transition model is exact, process noise would be zero. In real-world, it would be nearly impossible to capture the exact state-transition with a mathematical model. The process noise captures that uncertainty.
I need to make an application for creating logic circuits and seeing the results. This is primarily for use in A-Level (UK, 16-18 year olds generally) computing courses.
Ive never made any applications like this, so am not sure on the best design for storing the circuit and evaluating the results (at a resomable speed, say 100Hz on a 1.6Ghz single core computer).
Rather than have the circuit built from the basic gates (and, or, nand, etc) I want to allow these gates to be used to make "chips" which can then be used within other circuits (eg you might want to make a 8bit register chip, or a 16bit adder).
The problem is that the number of gates increases massively with such circuits, such that if the simulation worked on each individual gate it would have 1000's of gates to simulate, so I need to simplify these components that can be placed in a circuit so they can be simulated quickly.
I thought about generating a truth table for each component, then simulation could use a lookup table to find the outputs for a given input. The problem occurred to me though that the size of such tables increase massively with inputs. If a chip had 32 inputs, then the truth table needs 2^32 rows. This uses a massive amount of memory in many cases more than there is to use so isn't practical for non-trivial components, it also wont work with chips that can store their state (eg registers) since they cant be represented as a simply table of inputs and outputs.
I know I could just hardcode things like register chips, however since this is for educational purposes I want it so that people can make their own components as well as view and edit the implementations for standard ones. I considered allowing such components to be created and edited using code (eg dlls or a scripting language), so that an adder for example could be represented as "output = inputA + inputB" however that assumes that the students have done enough programming in the given language to be able to understand and write such plugins to mimic the results of their circuit which is likly to not be the case...
Is there some other way to take a boolean logic circuit and simplify it automatically so that the simulation can determine the outputs of a component quickly?
As for storing the components I was thinking of storing some kind of tree structure, such that each component is evaluated once all components that link to its inputs are evaluated.
eg consider: A.B + C
The simulator would first evaluate the AND gate, and then evaluate the OR gate using the output of the AND gate and C.
However it just occurred to me that in cases where the outputs link back round to the inputs, will cause a deadlock because there inputs will never all be evaluated...How can I overcome this, since the program can only evaluate one gate at a time?
Have you looked at Richard Bowles's simulator?
You're not the first person to want to build their own circuit simulator ;-).
My suggestion is to settle on a minimal set of primitives. When I began mine (which I plan to resume one of these days...) I had two primitives:
Source: zero inputs, one output that's always 1.
Transistor: two inputs A and B, one output that's A and not B.
Obviously I'm misusing the terminology a bit, not to mention neglecting the niceties of electronics. On the second point I recommend abstracting to wires that carry 1s and 0s like I did. I had a lot of fun drawing diagrams of gates and adders from these. When you can assemble them into circuits and draw a box round the set (with inputs and outputs) you can start building bigger things like multipliers.
If you want anything with loops you need to incorporate some kind of delay -- so each component needs to store the state of its outputs. On every cycle you update all the new states from the current states of the upstream components.
Edit Regarding your concerns on scalability, how about defaulting to the first principles method of simulating each component in terms of its state and upstream neighbours, but provide ways of optimising subcircuits:
If you have a subcircuit S with inputs A[m] with m < 8 (say, giving a maximum of 256 rows) and outputs B[n] and no loops, generate the truth table for S and use that. This could be done automatically for identified subcircuits (and reused if the subcircuit appears more than once) or by choice.
If you have a subcircuit with loops, you may still be able to generate a truth table. There are fixed-point finding methods which can help here.
If your subcircuit has delays (and they are significant to the enclosing circuit) the truth table can incorporate state columns. E.g. if the subcircuit has input A, inner state B, and output C, where C <- A and B, B <- A, the truth table could be:
A B | B C
0 0 | 0 0
0 1 | 0 0
1 0 | 1 0
1 1 | 1 1
If you have a subcircuit that the user asserts implements a particular known pattern such as "adder", provide an option for using a hard-coded implementation for updating that subcircuit instead of by simulating its inner parts.
When I made a circuit emulator (sadly, also incomplete and also unreleased), here's how I handled loops:
Each circuit element stores its boolean value
When an element "E0" changes its value, it notifies (via the observer pattern) all who depend on it
Each observing element evaluates its new value and does likewise
When the E0 change occurs, a level-1 list is kept of all elements affected. If an element already appears on this list, it gets remembered in a new level-2 list but doesn't continue to notify its observers. When the sequence which E0 began has stopped notifying new elements, the next queue level is handled. Ie: the sequence is followed and completed for the first element added to level-2, then the next added to level-2, etc. until all of level-x is complete, then you move to level-(x+1)
This is in no way complete. If you ever have multiple oscillators doing infinite loops, then no matter what order you take them in, one could prevent the other from ever getting its turn. My next goal was to alleviate this by limiting steps with clock-based sync'ing instead of cascading combinatorials, but I never got this far in my project.
You might want to take a look at the From Nand To Tetris in 12 steps course software. There is a video talking about it on youtube.
The course page is at: http://www1.idc.ac.il/tecs/
If you can disallow loops (outputs linking back to inputs), then you can significantly simplify the problem. In that case, for every input there will be exactly one definite output. Cycles however can make the output undecideable (or rather, constantly changing).
Evaluating a circuit without loops should be easy - just use the BFS algorithm with "junctions" (connections between logic gates) as the items in the list. Start off with all the inputs to all the gates in an "undefined" state. As soon as a gate has all inputs "defined" (either 1 or 0), calculate its output and add its output junctions to the BFS list. This way you only have to evaluate each gate and each junction once.
If there are loops, the same algorithm can be used, but the circuit can be built in such a way that it never comes to a "rest" and some junctions are always changing between 1 and 0.
OOps, actually, this algorithm can't be used in this case because the looped gates (and gates depending on them) would forever stay as "undefined".
You could introduce them to the concept of Karnaugh maps, which would help them simplify truth values for themselves.
You could hard code all the common ones. Then allow them to build their own out of the hard coded ones (which would include low level gates), which would be evaluated by evaluating each sub-component. Finally, if one of their "chips" has less than X inputs/outputs, you could "optimize" it into a lookup table. Maybe detect how common it is and only do this for the most used Y chips? This way you have a good speed/space tradeoff.
You could always JIT compile the circuits...
As I haven't really thought about it, I'm not really sure what approach I'd take.. but it would possibly be a hybrid method and I'd definitely hard code popular "chips" in too.
When I was playing around making a "digital circuit" simulation environment, I had each defined circuit (a basic gate, a mux, a demux and a couple of other primitives) associated with a transfer function (that is, a function that computes all outputs, based on the present inputs), an "agenda" structure (basically a linked list of "when to activate a specific transfer function), virtual wires and a global clock.
I arbitrarily set the wires to hard-modify the inputs whenever the output changed and the act of changing an input on any circuit to schedule a transfer function to be called after the gate delay. With this at hand, I could accommodate both clocked and unclocked circuit elements (a clocked element is set to have its transfer function run at "next clock transition, plus gate delay", any unclocked element just depends on the gate delay).
Never really got around to build a GUI for it, so I've never released the code.