I have a dataset with a column on which I can apply outlier removal logic of box-plot (all rows with value lower than (Q1 - 1.5 x IQR) and higher than lower than (Q3 + 1.5 x IQR). However, it is observed that after removing outliers, if the box-plot is plotted again and outliers are detected and removed using the same logic, some additional rows are also eliminated. This process can be repeated multiple times till no additional rows are removed even after applying this logic.
My queries are:
What does it mean when additional rows are removed when box-plot is made again and new outliers are found?
Should outliers be removed only once or should we repeat the process till there are no more removals?
Are there any recognized studies / papers / journal articles which have studied this?
Related
In my Anylogic model I have a population of agents (4 terminals) were trucks arrive at, are being served and depart from. The terminals have two parameters (numberOfGates and servicetime) which influence the departures per hour of trucks leaving the terminals. Now I want to tune these two parameters, so that the amount of departures per hour is closest to reality (I know the actual departures per hour). I already have two datasets within each terminal agent, one with de amount of departures per hour that I simulate, and one with the observedDepartures from the data.
I already compare these two datasets in plots for every terminal:
Now I want to create an optimization experiment to tune the numberOfGates and servicetime of the terminals so that the departure dataset is the closest to the observedDepartures dataset. Does anyone know how to do create a(n) (objective) function for this optimization experiment the easiest way?
When I add a variable diff that is updated every hour by abs( departures - observedDepartures) and put root.diff in the optimization experiment, it gives me the eq(null) is not allowed. Use isNull() instead error, in a line that reads the database for the observedDepartures (see last picture), but it works when I run the simulation normally, it only gives this error when running the optimization experiment (I don't know why).
You can use the absolute value of the sum of the differences for each replication. That is, create a variable that logs the | difference | for each hour, call it diff. Then in the optimization experiment, minimize the value of the sum of that variable. In fact this is close to a typical regression model's objectives. There they use a more complex objective function, by minimizing the sum of the square of the differences.
A Calibration experiment already does (in a more mathematically correct way) what you are trying to do, using the in-built difference function to calculate the 'area between two curves' (which is what the optimisation is trying to minimise). You don't need to calculate differences or anything yourself. (There are two variants of the function to compare either two Data Sets (your case) or a Data Set and a Table Function (useful if your empirical data is not at the same time points as your synthetic simulated data).)
In your case it (the objective function) will need to be a sum of the differences between the empirical and simulated datasets for the 4 terminals (or possibly a weighted sum if the fit for some terminals is considered more important than for others).
So your objective is something like
difference(root.terminals(0).departures, root.terminals(0).observedDepartures)
+ difference(root.terminals(1).departures, root.terminals(1).observedDepartures)
+ difference(root.terminals(2).departures, root.terminals(2).observedDepartures)
+ difference(root.terminals(3).departures, root.terminals(2).observedDepartures)
(It would be better to calculate this for an arbitrary population of terminals in a function but this is the 'raw shape' of the code.)
A Calibration experiment is actually just a wizard which creates an Optimization experiment set up in a particular way (with a UI and all settings/code already created for you), so you can just use that objective in your existing Optimization experiment (but it won't have a built-in useful UI like a Calibration experiment). This also means you can still set this up in the Personal Learning Edition too (which doesn't have the Calibration experiment).
The documentation of MALLET mentions following:
--num-iterations [NUMBER]
The number of sampling iterations should be a trade off between the time taken to complete sampling and the quality of the topic model.
MALLET provides furthermore an example:
// Run the model for 50 iterations and stop (this is for testing only,
// for real applications, use 1000 to 2000 iterations)
model.setNumIterations(50);
It is obvious that too few iterations lead to bad topic models.
However, does increasing the number of Gibbs sampling iterations necessarily benefit the quality of the topic model (measured by perplexity, topic coherence or on a downstream task)?
Or is it possible that the model quality decreases with the --num-iterations set to a too high value?
On a personal project, averaged over 10-fold cross-validation increasing the number of iterations from 100 to 1000 did not impact the average accuracy (measured as Mean Reciprocal Rank) for a downstream task. However, within the cross-validation splits the performance changed significantly, although the random seed was fixed and all other parameters kept the same. What part of background knowledge about Gibbs sampling am I missing to explain this behavior?
I am using a symmetric prior for alpha and beta without hyperparameter optimization and the parallelized LDA implementation provided by MALLET.
The 1000 iteration setting is designed to be a safe number for most collection sizes, and also to communicate "this is a large, round number, so don't think it's very precise". It's likely that smaller numbers will be fine. I once ran a model for 1000000 iterations, and fully half the token assignments never changed from the 1000 iteration model.
Could you be more specific about the cross validation results? Was it that different folds had different MRRs, which were individually stable over iteration counts? Or that individual fold MRRs varied by iteration count, but they balanced out in the overall mean? It's not unusual for different folds to have different "difficulty". Fixing the random seed also wouldn't make a difference if the data is different.
Background
I'm trying to convert an algorithm from sequential to parallel, but I am stuck.
Point and Figure Charts
I am creating point and figure charts.
Decreasing
While the stock is going down, add an O every time it breaks through the floor.
Increasing
While the stock is going up, add an X every time it breaks through the ceiling.
Reversal
If the stock reverses direction, but the change is less than a reversal threshold (3 units) do nothing. If the change is greater than the reversal threshold, start a new column (X or O)
Sequential vs Parallel
Sequentially, this is pretty straight forward. I keep a variable for the floor and ceiling. If the current price breaks through the floor or ceiling, or changes more than the reversal threshold, I can take the appropriate action.
My question is, is there a way to find these reversal point in parallel? I'm fairly new to thinking in parallel, so I'm sorry if this is trivial. I am trying to do this in CUDA, but I have been stuck for weeks. I have tried using the finite difference algorithms from NVidia. These produce local max / min but not the reversal points. Small fluctuations produce numerous relative max / min, but most of them are trivial because the change is not greater than the reversal size.
My question is, is there a way to find these reversal point in parallel?
one possible approach:
use thrust::unique to remove periods where the price is numerically constant
use thrust::adjacent_difference to produce 1st difference data
use thrust::adjacent_difference on 1st difference data to get the 2nd difference data, i.e the points where there is a change in the sign of the slope.
use these points of change in sign of slope to identify separate regions of data - build a key vector from these (e.g. with a prefix sum). This key vector segments the price data into "runs" where the price change is in a particular direction.
use thrust::exclusive_scan_by_key on the 1st difference data, to produce the net change of the run
Wherever the net change of the run exceeds a threshold, flag as a "reversal"
Your description of what constitutes a reversal may also be slightly unclear. The above method would not flag a reversal on certain data patterns that you might classify as a reversal. I suspect you are looking beyond a single run as I have defined it here. If that is the case, there may be a method to address that as well - with more steps.
I'm trying to create benchmarks for a variety of games that have 5 levels each. The goal is to train a model to convergence on 3 levels first, and then measure the learning curves on the remaining 2 levels.
Is there a general rule for how models should be trained on multiple levels? Should the training be done on one level after another?
Thanks very much for the help.
Suppose you are able to train for N levels in total (within the time constraints you may have).
I would not recommend the following setup:
Train N / 3 times on the first level
Train N / 3 times on the second level
Train N / 3 times on the second level
The risk with such a setup is that you first learn to play well on the first level, then forget everything you learned and "overfit" to the second level, and then forget again and overfit to the third level.
You'll want to make sure that you consistently keep a nice mix of levels throughout the entire training process, because the goal ultimately is to generalize and perform well on the (unseen) levels 4 and 5.
To do this, I'd recommend one of the following setups:
Train once on the first level
Train once on the second level
Train once on the third level
Repeat from step one again, untill you've trained the maximum N times
Alternatively:
Randomly select one of the first three levels to train.
Repeat until N times trained.
It may be possible to do even better with more sophisticated strategies. For example, you could try tracking your average performance per level over the last X times you've played a level, and prioritize levels in which you're not performing well yet (because apparantly you still have a lot to learn in those). This could, for instance, be done with a Multi-Armed Bandit strategy such as UCB1, where you use the negative recent performance as a "reward".
It may also be worth looking into the Learning track of the General Video Game AI competition (http://gvgai.net/). I believe that competition has precisely the setup you mentioned of three training levels plus two levels per game for evaluation (maybe this is even where your question came from?). You can have a look at what various participants in this competition are doing if their source code is available, and/or look up literature about the competition / competing entries.
Existing implementation:
In my implementation of Tic-Tac-Toe with minimax, I look for all boxes where I can get best result and chose 1 of them randomly, so that the same solution isn't displayed each time.
For ex. if the returned list is [1, 0 , 1, -1], at some point, I will randomly chose between the two highest values.
Question about Alpha-Beta Pruning:
Based on what I understood, when the algorithm finds that it is winning from one path, it would no longer need to look for other paths that might/ might not lead to a winning case.
So will this, like I feel, cause the earliest possible box that leads to the best solution to be displayed as the result and seem the same each time? For example at the time of first move, all moves lead to a draw. So will the 1st box be selected each time?
How can I bring randomness to the solution like with the minimax solution? One way that I thought about now could be to randomly pass the indices to the alpha-beta algorithm. So the result will be the first best solution in that randomly sorted list of positions.
Thanks in advance. If there is some literature on this, I'd be glad to read it.
If someone could post some good reference for aplha-beta pruning, That'll be excellent as I had a hard time understanding how to apply it.
To randomly pick among multiple best solutions (all equal) in alpha-beta pruning, you can modify your evaluation function to add a very small random number whenever you evaluate a game state. You should just make sure that the magnitude of that random number is never greater than the true difference between the evaluations of two states.
For example, if the true evaluation function for your game state can only return values -1, 0, and 1, you could add a randomly generated number in the range [0.0, 0.01] to the evaluation of every game state.
Without this, alpha-beta pruning doesn't necessarily find only one solution. Consider this example from wikipedia. In the middle, you see that two solutions with an evaluation of 6 were found, so it can find more than one. I do actually think it will still find all moves leading to optimal solutions at the root node, but not actually find all solutions deep down in the tree. Suppose, in the example image, that the pruned node with score of 9 in the middle actually had a score of 6. It would still get pruned there, so that particular solution wouldn't be found, but the move from root node leading to it (the middle move at root) would still be found. So, eventually, you would be able to reach it.
Some interesting notes:
This implementation would also work in minimax, and avoid the need to store a list of multiple (equally good) solutions
In more complex games than Tic Tac Toe, where you cannot search the complete state space, adding a small random number for the max player and deducting a small random number for the min player like this may actually slightly improve your heuristic evaluation function. The reason for this is as follows. Suppose in state A you have 5 moves available, and in state B you have 10 moves available, which all result in the same heuristic evaluation score. Intuitively, the successors of state B may be slightly better, because you had more moves available; in many games, having more moves available means that you are in a better position. Because you generated 10 random numbers for the 10 successors of state B, it is also a bit more likely that the highest generated random number is among those 10 (instead of the 5 numbers generated for successors of A)