Related
I am planning to use neural networks for approximating a value function in a reinforcement learning algorithm. I want to do that to introduce some generalization and flexibility on how I represent states and actions.
Now, it looks to me that neural networks are the right tool to do that, however I have limited visibility here since I am not an AI expert. In particular, it seems that neural networks are being replaced by other technologies these days, e.g. support vector machines, but I am unsure if this is a fashion matter or if there is some real limitation in neural networks that could doom my approach. Do you have any suggestion?
Thanks,
Tunnuz
It's true that neural networks are no longer in vogue, as they once were, but they're hardly dead. The general reason for them falling from favor was the rise of the Support Vector Machine, because they converge globally and require fewer parameter specifications.
However, SVMs are very burdensome to implement and don't naturally generalize to reinforcement learning like ANNs do (SVMs are primarily used for offline decision problems).
I'd suggest you stick to ANNs if your task seems suitable to one, as within the realm of reinforcement learning, ANNs are still at the forefront in performance.
Here's a great place to start; just check out the section titled "Temporal Difference Learning" as that's the standard way ANNs solve reinforcement learning problems.
One caveat though: the recent trend in machine learning is to use many diverse learning agents together via bagging or boosting. While I haven't seen this as much in reinforcement learning, I'm sure employing this strategy would still be much more powerful than an ANN alone. But unless you really need world class performance (this is what won the netflix competition), I'd steer clear of this extremely complex technique.
It seems to me that neural networks are kind of making a comeback. For example, this year there were a bunch of papers at ICML 2011 on neural networks. I would definitely not consider them abandonware. That being said, I would not use them for reinforcement learning.
Neural networks are a decent general way of approximating complex functions, but they are rarely the best choice for any specific learning task. They are difficult to design, slow to converge, and get stuck in local minima.
If you have no experience with neural networks, then you might be happier to you use a more straightforward method of generalizing RL, such as coarse coding.
Theoretically it has been proved that Neural Networks can approximate any function (given an infinite number of hidden neurons and the necessary inputs), so no I don't think the neural networks will ever be abandonwares.
SVM are great, but they cannot be used for all applications while Neural Networks can be used for any purpose.
Using neural networks in combination with reinforcement learning is standard and well-known, but be careful to plot and debug your neural network's convergence to check that it works correctly as neural networks are notoriously known to be hard to implement and learn correctly.
Be also very careful about the representation of the problem you give to your neural network (ie: the inputs nodes): could you, or could an expert, solve the problem given what you give as inputs to your net? Very often, people implementing neural networks don't give enough informations for the neural net to reason, this is not so uncommon, so be careful with that.
Do you know if anyone has tried to compile high level programming languages (java, c#, etc') into a recurrent neural network and then evolve them?
I mean that the whole process including memory usage is stored in a graph of a neural net, and I'm talking about complex programs (thinking about natural language processing problems).
When I say neural net I mean a directed weighted graphs that spreads activation, and the nodes are functions of their inputs (linear, sigmoid and multiplicative to keep it simple).
Furthermore, is that what people mean in genetic programming or is there a difference?
Neural networks are not particularly well suited for evolving programs; their strength tends to be in classification. If anyone has tried, I haven't heard about it (which considering I barely touch neural networks is not a surprise, but I am active in the general AI field at the moment).
The main reason why neural networks aren't useful for generating programs is that they basically represent a mathematical equation (numeric, rather than functional). Given some numeric input, you get a numeric output. It is difficult to interpret these in the context of a program any more complicated than simple arithmetic.
Genetic Programming traditionally uses Lisp, which is a pure functional language, and often programs are often shown as tree diagrams (which occasionally look similar to some neural network diagrams - is this the source of your confusion?). The programs are evolved by exchanging entire branches of a tree (a function and all its parameters) between programs or regenerating an entire branch randomly.
There are certainly a lot of good (and a lot of bad) references on both of these topics out there - I refrain from listing them because it isn't clear what you are actually interested in. Wikipedia covers each of these techniques, and is a good starting point.
Genetic programming is very different from Neural networks. What you are suggesting is more along the lines of genetic programming - making small random changes to a program, possibly "breeding" successful programs. It is not easy, and I have my doubts that it can be done successfully across a large program.
You may have more luck extracting a small but critical part of your program, one which has a few particular "aspects" (such as parameter values) that you can try to evolve.
Google is your friend.
Some sophisticated anti-virus programs as well as sophisticated malware use formal grammar and genetic operators to evolve against each other using neural networks.
Here is an example paper on the topic: http://nexginrc.org/nexginrcAdmin/PublicationsFiles/raid09-sadia.pdf
Sources: A class on Artificial Intelligence I took a couple years ago.
With regards to your main question, no one has ever tried that on programming languages to the best of my knowledge, but there is some research in the field of evolutionary computation that could be compared to something like that (but it's obviously a far-fetched comparison). As a matter of possible interest, I asked a similar question about sel-improving compilers a while ago.
For a difference between genetic algorithms and genetic programming, have a look at this question.
Neural networks have nothing to do with genetic algorithms or genetic programming, but you can obviously use either to evolve neural nets (as any other thing for that matters).
You could have look at genetic-programming.org where they claim that they have found some near human competitive results produced by genetic programming.
I have not heard of self-evolving and self-imrpvoing programs before. They may exist as special research tools like genetic-programming.org have but nothing solid for generic use. And even if they exist they are very limited to special purpose operations like malware detection as Alain mentioned.
Why do we use neural networks? It's biologic. Aren't there any more solutions that're more "suitable" for computers?
In other words: Why do we use the human brain as a model for inspiration for artifical intelligence?
Neural networks aren't really very biological. They resemble, at a very general level, the architecture of neurons, but it's a great exaggeration to say that they work "just like the brain" (an exaggeration that's encouraged by some neural-net advocates, alas).
Neural nets are mostly used for fuzzy, difficult problems that don't yield to traditional algorithmic approaches. IOWs, there are more "suitable" solutions for computers, but sometimes those solutions don't work, and in those cases one approach is a neural network.
Why do we use neural networks?
Because they're simple to construct, and often appear to be a good approach to certain classes of problems, such as pattern recognition.
Aren't there any more solutions that're more "suitable" for computers?
Yes, implementations that more closely match a computer's architecture can be more suitable for the computer, but then can be less suitable for an effective solution.
Why do we use the human brain as a model for inspiration for artifical intelligence?
Because our brain is the superior example we have of something intelligent.
Neural Networks are still used for two reasons.
They are easy to understand for people who don't want to delve into the math of a more complicated algorithm.
They have a really good name. I mean when you role into a CEO's office to sell him your model which would you rather say, Neural Network or Support Vector Machine. When he asks how it works you can just say "just like the neurons in your brain", which is something most people understand. If you try and explain a support vector machine Mr. CEO is going to be lost (Not because he is dumb but because SVMs are harder to understand).
Sometimes they are still useful however I think that the training time is often just too long.
I don't understand the question. Neural nets are suitable for certain functions, and not others. The same is true for various other sorts of classes of algorithms, regardless of what they might have been inspired by.
If we have a good many inputs to something, and we want some outputs, and we have a set of example inputs with known desired outputs, and we don't want to calculate a function ourselves, neural nets are excellent. We feed in the example inputs, compare the output to the example outputs, and adjust the inner workings of the NN in an automatic fashion, to make the NN output closer to the desired output.
This sort of function derivation is very useful in various forms of pattern recognition and general classification. It isn't a panacea, of course. It has no explanatory power (in that you can't look at the innards to see why it classifies something in a particular way), it doesn't offer guarantees of correctness within certain limits, validating how well it works is difficult, and gathering enough examples for training and validation can be expensive or even impossible. The trick is to know when to use a NN and what sort to use.
There are, of course, people who oversell the things as some sort of super solution or even an explanation of human thought, and you might be reacting to them.
Neural network are only "inspired" by the neural structure of our brain, but they are not even close to the complexity of the behaviour of a real neuron (to date there is no neuron model that captures the complexity of a SINGLE neuron, don't even think about a neuronal population...)
Although "neural", machine "learning" and other "pseudo-bio" (like "genetic algorithms") terms are very "cool", that does not mean that they are actually based on real biological processes.
Just that they may very approximatively remind of a biological situation.
NB: of course this does not make them useless! They're very very important in many fields!
Neural networks have been around for a while, and originally were developed to model as close an understanding as we had at the time to the way neurons work in the brain. They represent a network of neurons, hence "neural network." Since computers and brains are very different hardware-wise, implementing anything like a brain with a computer is going to be rather clunky. However, as others have stated so far, neural networks can be useful for some things that are vague such as pattern recognition, facial recognition, and other similar uses. They are also still useful as a basic model of how neurons connect and are often used in Cognitive Science and other fields of artificial intelligence to try to understand how small parts of the complex human brain might make simple decisions. Unfortunately, once a neural network "learns" something, it is very difficult to understand how it actually makes its decisions.
There are, of course, many misuses of neural networks and in most non-research applications, other algorithms have been developed that are much more accurate. If a piece of business software proudly proclaims it uses a neural network, chances are it probably doesn't need it, and might be using it to inefficiently perform a task that could be performed in a much easier way. Unless the software is actually "learning" on the fly, which is very rare, neural networks are pretty much useless. And even when the software is "learning", sometimes neural networks aren't the best way to go.
While I admit, I tinker with Neural Networks because of my hopes in creating high level AI, however, you can look at a Neural Network as being more than just just an artificial representation of a human brain, but as a Mathematical construct.
For example Let's say you have a function y = f(x) or more abstractly y = f(x1, x2, ..., xn-1, xn), Neural networks themselves act as functions, or even a set of functions, taking in a large input and producing some output [y1, y2, ..., yn-1, yn] = f(x1, x2, ..., xn-1, xn)
Furthermore, they are not static, but instead can continue adapting and learning and eventually extrapolate(predict) interesting things. Their abstractness can even result in them coming up with unique solutions to problems that haven't haven't been thought up yet. For example the TDGammon program learned to play backgammon and beat the world champion. The world champion stated that the program play a unique end game that he had never seen. (that's pretty awesome if you ask me considering the complexity of NNs)
And then when you look at recurrent neural networks (i.e. can have internal feedback loops, or pipe their output back into their input, while consuming new input) they can solve even more interesting problems, and map even more complex functions.
In a nutshell Neural Networks are like a very very abstract high dimensional function and capable of mapping/learning very interesting things that would be otherwise impossible to program programmatically. For example, the energy needed to calculate the total net Forces of Gravity on a large number of objects is intense (you have to calculate it for each object, and against each object), but once a neural network learns how to map it they can do these complex calculations that would run in exponential or combinatoric? time in polynomial time. Just look at how fast your brain processes physics data, spatial data/ images / sound when you dream. That's the potential computation power of Neural Networks. And to also mention the way they store data is very clever as well (in synaptics patterns, i.e. memories)
Artificial intelligence is a branch of computer science devoted to making computers more 'biologic.' This is useful when you want a computer to do human(biologic) things like play chess, or imitate casual conversation.
Human brains are much more efficient and powerful in some ways than the most powerful computers, so it makes sense to try to imitate a biological way of processing information.
Most neural networks I'm aware of are nothing more than flexible interpolators. Backpropagating of errors is easy and fast, here are some possible uses :
Classification of data
Some games (modern backgammon AIs beat the best players in the world, the evaluation function is a neural net)
Pattern recognition (OCR ?)
There is nothing particularly related to human intelligence. There are other uses of neural nets, I have seen an implementation of associative memory which allowed for degradation without (much) data loss, pretty much like the brain which sees some neurons die with time.
Seeing that as as far as we know, one half of your brain is logical and the other half of your brain is emotional, and that the wants of the emotional side are fed to the logical side in order to fulfill those wants; has there been any research done in connecting two separate neural networks to one another (one trained to be emotional, and one trained to be logical) to see if it would result in almost a free-will sort of "brain"?
I don't really know anything about neural networks except that they were modeled after the biological synapses in the human brain, which is why I ask.
I'm not even sure if this would be possible considering that even a trained neural network sometimes doesn't act logically (a.k.a. do what you thought you trained it to do).
First, most modern neural networks aren't really modeled after biological synapses. They use an Artificial Neuron which allowed Back Propagation to work rather than a Perceptron which is a much more accurate representation.
When you feed the output of one network into the input of another network, you've really just created one larger network, not two separate networks. It just happens that in this case portions of the networks would be trained independently.
That said, all neural networks have to be trained. Which means you need sample input and sample output. You are looking to create a decision engine of sorts I suppose. So you would need to create a dataset where it makes sense that there might be an emotional and rational response, such as purchasing an item. You'd have to train the 'rational' network to accept as a set of inputs the output of an 'emotional' network. Which means you are really just training the rational decision engine to be responsive based on the leve of 'distress' caused by the emotional network.
Just my two cents.
I have also heard of one hemisphere being called "divergent" and one "convergent". This may not make any more sense than emotional vs logical, but it does hint at how you might model it more easily. I don't know how the brain achieves some of the impressive computational feats it does, but I wouldn't be very surprised if all revolved around balance, but maybe that is just one of the baises you have when you are a brain with two hemipheres (or any even number) :D
A balance between convergence and divergence is the crux of the creativity inherent in evolution. Replicating this with neural nets sounds promising to me. Suppose you make one learning system that generalizes and keeps representations of only the typical groups of patterns it is shown. Then you take another and make it generate all the in-betweens and mutants of the patterns it is shown. Then you feed them to eachother in a circle, and poof, you have made something really interesting!
It's even more complex than that, unbelievably. The left hemisphere works on a set of logical rules, it uses these to predict its environment and categorize input. It also infers rules and stores them for future use. The right hemisphere is based, as you said, on emotion, but also on memory of single, unique or emotionally relevant occurrences. A software implementation should also be able to retrieve and store these two data types and exchange "opinions" about them.
While the left hemisphere of the brain may be more involved in making emotional decisions, emotion itself is unlikely to occur exclusively in one side of the brain, and the interplay between emotions and rational thought within the brain is likely to be substantially more complex than having two completely separate circuits. For instance, a study on rhesus macaques found that dopamine and other hormones associated with emotional responses essentially implements temporal difference learning within the brain (I'm still looking for a link to it). This suggests that separating emotional and rational thought into two separate neural networks probably wouldn't be practical, even if we had the resources to build neural networks on the scale of brain hemispheres (which we don't, or at least not within most research budgets).
This idea is supported by Sloman and Croucher's suggestion that emotion will likely be an unavoidable emergent property of a sufficiently advanced intelligent system. Such systems (discussed in detail in the paper) will be much more complex than straight-up neural nets. More importantly, though, the emotions won't be something that you can localize to one part of the system.
Is a genetic algorithm the most efficient way to optimize the number of hidden nodes and the amount of training done on an artificial neural network?
I am coding neural networks using the NNToolbox in Matlab. I am open to any other suggestions of optimization techniques, but I'm most familiar with GA's.
Actually, there are multiple things that you can optimize using GA regarding NN.
You can optimize the structure (number of nodes, layers, activation function etc.).
You can also train using GA, that means setting the weights.
Genetic algorithms will never be the most efficient, but they usually used when you have little clue as to what numbers to use.
For training, you can use other algorithms including backpropagation, nelder-mead etc..
You said you wanted to optimize number hidden nodes, for this, genetic algorithm may be sufficient, although far from "optimal". The space you are searching is probably too small to use genetic algorithms, but they can still work and afaik, they are already implemented in matlab, so no biggie.
What do you mean by optimizing amount of training done? If you mean number of epochs, then that's fine, just remember that training is somehow dependent on starting weights and they are usually random, so the fitness function used for GA won't really be a function.
A good example of neural networks and genetic programming is the NEAT architecture (Neuro-Evolution of Augmenting Topologies). This is a genetic algorithm that finds an optimal topology. It's also known to be good at keeping the number of hidden nodes down.
They also made a game using this called Nero. Quite unique and very amazing tangible results.
Dr. Stanley's homepage:
http://www.cs.ucf.edu/~kstanley/
Here you'll find just about everything NEAT related as he is the one who invented it.
Genetic algorithms can be usefully applied to optimising neural networks, but you have to think a little about what you want to do.
Most "classic" NN training algorithms, such as Back-Propagation, only optimise the weights of the neurons. Genetic algorithms can optimise the weights, but this will typically be inefficient. However, as you were asking, they can optimise the topology of the network and also the parameters for your training algorithm. You'll have to be especially wary of creating networks that are "over-trained" though.
One further technique with a modified genetic algorithms can be useful for overcoming a problem with Back-Propagation. Back-Propagation usually finds local minima, but it finds them accurately and rapidly. Combining a Genetic Algorithm with Back-Propagation, e.g., in a Lamarckian GA, gives the advantages of both. This technique is briefly described during the GAUL tutorial
It is sometimes useful to use a genetic algorithm to train a neural network when your objective function isn't continuous.
I'm not sure whether you should use a genetic algorithm for this.
I suppose the initial solution population for your genetic algorithm would consist of training sets for your neural network (given a specific training method). Usually the initial solution population consists of random solutions to your problem. However, random training sets would not really train your neural network.
The evaluation algorithm for your genetic algorithm would be a weighed average of the amount of training needed, the quality of the neural network in solving a specific problem and the numer of hidden nodes.
So, if you run this, you would get the training set that delivered the best result in terms of neural network quality (= training time, number hidden nodes, problem solving capabilities of the network).
Or are you considering an entirely different approach?
I'm not entirely sure what kind of problem you're working with, but GA sounds like a little bit of overkill here. Depending on the range of parameters you're working with, an exhaustive (or otherwise unintelligent) search may work. Try plotting your NN's performance with respect to number of hidden nodes for a first few values, starting small and jumping by larger and larger increments. In my experience, many NNs plateau in performance surprisingly early; you may be able to get a good picture of what range of hidden node numbers makes the most sense.
The same is often true for NNs' training iterations. More training helps networks up to a point, but soon ceases to have much effect.
In the majority of cases, these NN parameters don't affect performance in a very complex way. Generally, increasing them increases performance for a while but then diminishing returns kick in. GA is not really necessary to find a good value on this kind of simple curve; if the number of hidden nodes (or training iterations) really does cause the performance to fluctuate in a complicated way, then metaheuristics like GA may be apt. But give the brute-force approach a try before taking that route.
I would tend to say that genetic algorithms is a good idea since you can start with a minimal solution and grow the number of neurons. It is very likely that the "quality function" for which you want to find the optimal point is smooth and has only few bumps.
If you have to find this optimal NN frequently I would recommend using optimization algorithms and in your case quasi newton as described in numerical recipes which is optimal for problems where the function is expensive to evaluate.