I am attempting to train a neural network to control a simple entity in a simulated 2D environment, currently by using a genetic algorithm.
Perhaps due to lack of familiarity with the correct terms, my searches have not yielded much information on how to treat fitness and training in cases where all the following conditions hold:
There is no data available on correct outputs for given inputs.
A performance evaluation can only be made after an extended period of interaction with the environment (with continuous controller input/output invocation).
There is randomness inherent in the system.
Currently my approach is as follows:
The NN inputs are instantaneous sensor readings of the entity and environment state.
The outputs are instantaneous activation levels of its effectors, for example, a level of thrust for an actuator.
I generate a performance value by running the simulation for a given NN controller, either for a preset period of simulation time, or until some system state is reached. The performance value is then assigned as appropriate based on observations of behaviour/final state.
To prevent over-fitting, I repeat the above a number of times with different random generator seeds for the system, and assign a fitness using some metric such as average/lowest performance value.
This is done for every individual at every generation. Within a given generation, for fairness each individual will use the same set of random seeds.
I have a couple of questions.
Is this a reasonable, standard approach to take for such a problem? Unsurprisingly it all adds up to a very computationally expensive process. I'm wondering if there are any methods to avoid having to rerun a simulation from scratch every time I produce a fitness value.
As stated, the same set of random seeds is used for the simulations for each individual in a generation. From one generation to the next, should this set remain static, or should it be different? My instinct was to use different seeds each generation to further avoid over-fitting, and that doing so would not have an adverse effect on the selective force. However, from my results, I'm unsure about this.
It is a reasonable approach, but genetic algorithms are not known for being very fast/efficient. Try hillclimbing and see if that is any faster. There are numerous other optimization methods, but nothing is great if you assume the function is a black box that you can only sample from. Reinforcement learning might work.
Using random seeds should prevent overfitting, but may not be necessary depending on how representative a static test is of average, and how easy it is to overfit.
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 have run ANN in matlab for prediction a variable based on several response variables.ALL variables have numerical values.I could not get a desirable results although I changed hidden neuron several times many runs of the model and so on.My question is should I use transformation of the input variables to get a better results?how can I know that which transformation I should choos?Thanks for any help.
I strongly advise you to use some methods from time series analysis like lagged correlation or window lagged correlation (with statistical tests). You can find it in most of statistical packages (e.g. in R). From one small picture it's hard to deduce whether your prediction is lagged or not. Testing huge amount of data can help you in revealing true dependencies and avoid trusting in spurious correlations.
I have used caliper based template https://github.com/dcsobral/scala-foreach-benchmark happily for years. It runs randomly constructed problems multiple times and than calculates average time consumption.
Now I have faced with non-deterministic algorithm. So I need to know both running time and resulting fitness. I'm searching for a java/scala benchmark framework that can measure both characteristics in average and worst case scenario.
Non-deterministic means that algorithm relies on some random generator to make decisions. It used to find a near-optimal solution, where searching the optimum would require too much processor time. e.g. solution for the TSP problem.
Fitness means cost function for optimization process. Different runs (with different random seeds) may come with different costs. So you need to stabilize not only run time, but cost value (fitness).
I don't know if repeated invocation of a function until it shows acceptable run time variation is common feature of benchmark frameworks, but the caliper does so, and I search for a similar framework that is more advanced and capable of handling fitness besides time.
Use JMH: it has Scala bindings, and can sample execution time.
I am trying to solve classification problem using Matlab GPTIPS framework.
I managed to build reasonable data representation and fitness function so far and got an average accuracy per class near 65%.
What I need now is some help with two difficulties:
My data is biased. Basically I am solving binary classification problem and only 20% of data belongs to class 1, while other 80% belong to class 0. I used accuracy of prediction as my fitness function at first, but it was really bad. The best I have now is
Fitness = 0.5*(PositivePredictiveValue + NegativePredictiveValue) - const*ComplexityOfSolution
Please, advize, how can I improve my function to make correction for data bias.
Second problem is overfitting. I divided my data into three parts: training (70%), testing (20%), validation (10%). I train each chromosome on training set, then evaluate it's fitness function on testing set. This routine allows me to reach fitness of 0.82 on my test data for the best individual in population. But same individual's result on validation data is only 60%.
I added validation check for best individual each time before new population is generated. Then I compare fitness on validation set with fitness on test set. If difference is more then 5%, then I increase penalty for solution complexity in my fitness function. But it didn't help.
I could also try to evaluate all individuals with validation set during each generation, and simply remove overfitted ones. But then I don't see any difference between my test and validation data. What else can be done here?
UPDATE:
For my second question I've found great article "Experiments on Controlling Overtting
in Genetic Programming" Along with some article authors' ideas on dealing with overfitting in GP it has impressive review with a lot of references to many different approaches to the issue. Now I have a lot of new ideas I can try for my problem.
Unfortunately, still cant' find anything on selecting a proper fitness function which will take into account unbalanced class proportions in my data.
65% accuracy is very bad when the baseline (classify everything as the class with most samples) would be 80%. You need to achieve at least baseline classification in order to have a better model than the naive one.
I would not penalize complexity. Rather limit the tree size (if possible). You could identify simpler models during the run, like storing a pareto front of models with quality and complexity as its two fitness values.
In HeuristicLab we have integrated GP based classification that can do these things. There are several options: You can choose to use MSE for classification or R2. In the latest trunk build there is also an evaluator to optimize accuracy directly (exactly speaking it optimizes the classification penalties). Optimizing MSE means it assigns each class a value (1, 2, 3,...) and tries to minimize mean squared error from that value. This may not seem optimal at first, but works. Optimizing accuracy directly may lead to faster overfitting. There is also a formula simplifier which allows you to prune and shrink your formula (and view the effects of that).
Also, does it need to be GP? Have you tried Random Forest Classification or Support Vector Machines as well? RF are pretty fast and work pretty well usually.
Beginner on ANNs:
I am implementing a back propagation neural network to predict the price of gold. I know that I have to split my data into training data, selection data and test data.
However I unsure How to go on about using these sets of data. At first I was training the data network with my training set then after it's trained I am getting a number of inputs to my network from the test set and comparing the output.
I'm not sure if I'm doing this right and were does the selection set come in ?
thanks in advance!
The general idea is:
Train the network for a little while on the training set.
Evaluate the network on a second set, often called the validation set. Probably what you're calling the selection set.
Train the network a little more on the training set.
Evaluate the new network on the selection set again.
Which did better, the old network or the new network? If the new network is better, we're still getting some use out of training, so goto 3. If the new network is worse, more training will probably only hurt. Use the previously version of the network, since it did better.
In this way, you can tell when to stop training.
One easy modification to this is to always keep track of the best network seen so far, and we only stop training when we see some number (say, three) of training attempts that do worse in a row.
The third set, the test set, is necessary because the selection set is, if indirectly, involved in the training process. Final evaluation must be done on data that was not used at all during training.
This sort of thing is sufficient for simple experiments, but in general you'll want to use cross-validation to get a better idea of your system's performance.
I wanted to leave a comment just to say that validation sets are a good place for model-dependent hyper-parameter tuning, but I'm new here and hence lack the reputation points to do so. To make this more worthy of a separate posting, I've included an outline of my own train-validate-test process. In practice, my workflow is as follows:
Identify, collect, and clean data. Try to limit complaining during data munging process.
Split data into three sets: training, validation, test.
Establish two "base" models for evaluating more complex models built later on in the process. The first of these models is typically a basic linear/logistic regression using all possible features. The second models uses only the most obviously informative (initial identification of informative features depends on use case, typically involves combination of domain knowledge, basic clustering, simple correlation).
Begin more empirical feature selection (i.e. unsupervised NN, but usually random forest) and prototype a broad range of models using the training set.
Eliminate poorly performing models as well as uninformative features
Compare performance of remaining models against each other and the "base" models, using a modified version of the training set (same data, but sans uninformative features). Toss under-performing models.
Using the validation set, tune the appropriate hyper-parameters for each of the models (either by hand or gridsearch). Further reduce the number of models in consideration, ideally to just 2-3 (excluding base models).
Finally, evaluate model performance (with optimized hyper-parameters) on the test set. Again, compare models among themselves and against the base models. Make final model choice based on a problem-specific appropriate combination of computational complexity/cost, ease of interpretation/transparency/"explainability", and improvement over and/or performance vs base models.