How neural nets inspire communication patterns between intelligent agents
A communication plays major roles in a human society. It takes a long time for a newborn to learn a basic language, a means of communication. However, even a lifetime is short to fully master an art of communication. Recent trends in artificial intelligence, machine learning and robotics might tell us that we are about to welcome a new kind of intelligent agent, namely artificial intelligent agents or bots for short. The question is how are they going to communicate? If they were to combine to perform some task, they need to be able to form a quite effective communication patterns which facilitates an accomplishment of the given task. One possibility is for us to instruct robots on effective ways of communication. However, humans cannot do so all the time, otherwise it would defeat the purpose of having bots in the first place. So, our bots need to learn how to communicate at some point. This question bothered me for some time, not that I am a bot, I am just curious. I do not know why, but I just can’t see how these bots could learn complex communication patterns. The more I thought about the very nature of communication, the more I felt perplexed. After all, what is a communication? So I decided to supply my own definition of a communication. I defined it as:
A communication is an exchange of data with other agents directed towards the fulfillment of a particular goal
Though this definition falls short in many aspects, it seems to be pretty reasonable. Take, for instance, a communication of ants using pheromonones, whereby ants inform each other of paths towards a certain food source, which in turn improves well-being of their colony. Having identified my definition of communication, I set out to look for examples of communication both in natural and man-made worlds. There are fascinating examples of communication in nature. For instance, a bee’s dance is found to encode a direction towards flowers. This could be one source of inspiration. Another sources of inspirations are found among man-made objects. One advantage of such objects is the fact they are much better understood compared to natural examples. Among man-made examples of displaying wonders of communication patterns, I would vote for neural networks. In fact, these guys can be viewed as an epitome of communication. A node in any neural network is as trivial as it could get, boiling down to one simple operations of addition, nonlinearity operator and few others. However, when thousands or millions of these guys come together miracles of a pattern recognition start to happen.
It is possible to learn basic patterns of communication from neural nets. Consider a task of regression using neural networks. Say we have a dataset of D of certain size, where input and output are of dimensions n and m respectively. Our objective is to approximate a map f from an input space to an output space provided the dataset. One way to look at this problem is via multi agent learning. Suppose n bots are assumed to be input receivers, while m bots are assigned a job of output producers. All other bots are placed as nodes of some neural network architecture. There are two different kinds roles these bots can serve:
- Summation
- Nonlinear units, e.g. ReLU
The network structure in above example can be viewed as a communication architecture and an instance of weights can be viewed as a communication pattern. An objective is to reach an optimal communication pattern which ensures best regression results. In the case of neural networks, this objective can be achieved through gradient descent using backpropagation algorithm.
This post tried to shift perspectives of looking at multi-agent communication problem, where common neural network training can serve as an instance of such a problem. Certainly, a problem of neural network training is much simpler than that of multi-agent learning. Hence, above transformation does not constitute a proper reduction. Nevertheless, it might be beneficial to look at problems from different, sometimes unusual angles. Let me know what do you think about this. Thanks.
