Today we’re introducing Neuro, our new manager of coin emission, rewards and mining difficulty. Neuro is a neural network, a specialised AI, embedded in every Enecuum client.It actively adapts to market circumstances and predicts the optimal network parameters for all clients.
Introduction and background
In this section we’ll go over the basics of blockchain, crypto currency and neural networks, feel free to skip to the next chapter if you’re already familiar with this.
Blockchain and cryptocurrency
This article focuses on the rewards for — and difficulty of completing network consolidation tasks: Mining (Proof of Work), Verification (Proof of Activity) and Stake.We’ll go over the different task types later.
A traditional blockchain uses a single type of task, normally a Proof of Work, to mine new blocks. Each time this task is completed others verify the work and a new task is started,we will call this one cycle. The reward for completing a task is generally fixed by a preprogrammed formula called the emission curve that takes the mining cycle number and outputs the reward for that specific cycle. Once a block is mined, verified and accepted by the clients the cycle starts over.
The average time it takes to complete a mining task is determined by the performance of all clients and its diffilculty level. The diffculty is, in most blockchains, determined by an algorithm that attempts to maintain a target completion time, for example 2,5 or 10 minutes per block. In practice the completion times of traditional networks are rather volatile.
Bitcoin block completion times, in minutes, over 1 week (September 2018).
Blockchains are, amongst other things, distributed networks. To be fully distributed a few conditions need to be met. Two that are relevant to Enecuum’s AI are:
- Each client needs to be able to independently derive all information from the blockchain for the network to be completely distributed. For example: If a client needs to contact a server or other client to obtain the reward for mining the next block then the network is partly centralised.
- Each client needs to process the information in a blockchain in exactly the same manner every time, without external verification required. This is called deterministic behaviour. This is required to be able to reach a consensus.
Both conditions increased the challenge of integrating machine learning systems in a blockchain ecosystem.
Machine learning and neural networks
Like the building blocks for blockchains, the foundation for artificial neural networks have been around for a long time, at least since the 1940s. And, again like blockchains, improvement both to algorithms and technology in recent years have resulted in this technology gaining critical momentum; the AI is here to stay.
An artificial neural network is loosely based on the workings of a human mind. As our understanding of how the mind operates on a most fundamental level increased, neural networks diverged and re-converged multiple times during their history.
The above image shows a neuron cell, a single unit of the vast network that forms our mind. For our purposes the most important parts are the Cell body and Synaptic terminals, machine learning borrows these terms and most types of neural network attempt to emulate these specific parts, the cell body as a small data processing unit, the synaptic terminals as communication to cells in the next network layer. In a biological mind the connecting cells (in a cerebral cortex) look like this:
In artificial neural networks this is generally represented by a layered network, where each neuron cell body is connected to some or all neurons in the ‘next’ layer. Like biological synapses, each connection has a different rate of passing along a signal, called a weight. In biological neural networks a ‘neuron’ refers to the whole unit pictured, including synapses. In artificial networks a neuron normally refers to the cell body only.
Information in a layered network flows from left to right, from input neurons to, eventually, output neurons. This allows for relatively easy and precise training by a method called back-propagation. Other structures and methods exist, but are sofar less successful.
Let’s look at a simple network with two input neurons, one hidden neuron and one output neuron:
This very simple model can be reduced to a linear equation, which means that no matter how many neurons are present, only problems that behave linear can be solved. To enable a neural network to solve more complex problems a source of non-linearity, tanh, is added:
Now the state of the hidden₁ neuron is no longer linearly related to the input. There are other sources of non-linearity as well, each with different trade-offs. The computing done ‘inside’ a neuron can be expanded to do more than just introduce non-linear behaviour, more on that later.
The mining market
Mining a blockchain network, like Enecuum’s, is a growing and open market system with strong external influences (marketing, new technologies, economic and legislative changes, to name a few). Because of this it is practically impossible to create an algorithm to perfectly predict the required diffculties and rewards. In Enecuum’s case Proof of Work, Proof of Activity and Proof of Stake jobs all have partial interaction and competition with each other, making our market interactions an order of a magnitude more complex than the simpler single product market a Proof of Work-only blockchain. Such a blockchain normally uses a variation on a time-series algorithm, ranging from moving averages to specialised algorithms resembling, for example, (S)ARIMA.
Doing away with fixed algorithms allows for more flexibility and market-oriented behaviour. For example, in Enecuum it’s no longer required to have a fixed emission scheme; the emission behaviour can be tuned to the actual network requirements.
The second article is available by this link:
Join Enecuum socials in order to be instantly informed about the project updates: