superhuman engines on common hardware?

[smatovic]

I say the race for super human engines just started.

So, when will we see top engines based on neural networks
running on common hardware?

--
Srdja

[Milos]

I say the race for super human engines just started.

So, when will we see top engines based on neural networks
running on common hardware?

--
Srdja

How does 5-10 years qualify, matter of years or matter of decades?

[Dirt]

Years.

[Daniel Shawul]

There is not a lot of expertise here in either NN or MCTS, as opposed to computer Go people so I would say years. After AlphaGo defeated Lee and published their methods many former top engines adopted similar approach in months on commodity GPUs (not TPUs) to improve greatly. Stockfish basically needs to written from scratch to adopt AlphaZero's method and it would be an entirely new project anyway for them. Throw away alpha-beta, null-move, lmr, see, qsearch etcc... whats left of the past 50 years of chess research?

I think it would be worthwhile to program deeplearning software anyway, as you might find it useful in your career especially if you are a software engineer.

Daniel

P.S: NebiyuAlien can already play chess & shogi with MCTS (though very crapily) and a material eval. The MTCS was awesome for Checkers though! What I need to do is replace the material eval with deep NNs in a few months

[jhellis3]

Money no object: 2018.

Commonly available for enthusiasts with a reasonable budget: 2 - 3 years.

Ubiquitous: 4 - 5+ years.

[duncan]

I say the race for super human engines just started.

So, when will we see top engines based on neural networks
running on common hardware?

--
Srdja

I do not understand how this software works.? is it just software that has found a better evaluation that can run on a regular computer with no search?

[smatovic]

I do not understand how this software works.? is it just software that has found a better evaluation that can run on a regular computer with no search?

...i will give it a try...

The evaluation knowledge is stored in an neural network,
and is obtained by selfplay,
instead of classic, handcrafted and tuned evaluation terms.

Deepmind decided to start with zero knowledge in this network,
only the rules of chess were applied,
so AlphaZero it is able to discover unknown coherences
and find moves which are out of human bound.

To train that network Deepmind used 5000+64 TPUs,
to query that network during the match they used 4 TPUs.

How such an design can be ported to common hardware is matter of discussion.

So, yes, in the optimal case, AlphaZero would perform only a depth 1 search and already find the best move.

But in front of the neural network they perform an selective Monte Carlo Tree Search
with 80 Knps (single node, 4 TPUs).

--
Srdja

[duncan]

I do not understand how this software works.? is it just software that has found a better evaluation that can run on a regular computer with no search?

...i will give it a try...

The evaluation knowledge is stored in an neural network,
and is obtained by selfplay,
instead of classic, handcrafted and tuned evaluation terms.

Deepmind decided to start with zero knowledge in this network,
only the rules of chess were applied,
so AlphaZero it is able to discover unknown coherences
and find moves which are out of human bound.

To train that network Deepmind used 5000+64 TPUs,
to query that network during the match they used 4 TPUs.

How such an design can be ported to common hardware is matter of discussion.

So, yes, in the optimal case, AlphaZero would perform only a depth 1 search and already find the best move.

But in front of the neural network they perform an selective Monte Carlo Tree Search
with 80 Knps (single node, 4 TPUs).

--
Srdja

thanks for your reply. you say the evaluation knowledge is stored in a neural network. what is the actual neural network ? is it the actual 5000+64 TPUs or something else? and can it not be written into code and put in stockfish evaluation's function?

[duncan]

Money no object: 2018.

about how much money today ?

[smatovic]

thanks for your reply. you say the evaluation knowledge is stored in a neural network. what is the actual neural network ? is it the actual 5000+64 TPUs or something else?

i am just a laymen in this topic, anyway:

A neural network consists of knots and edges and resists in memory.

A layer of knots is connected via weighted edges with other layers of knots.

Each knot or layer can represent a feature to map.

NeuroChess for example had 175 handcrafted features.

http://chessprogramming.wikispaces.com/NeuroChess

During training the weights of the edges are adjusted,
and the network learns which combination of features are 'good' or 'bad'.

and can it not be written into code and put in stockfish evaluation's function?

Point is that current neural networks are black boxes,
you can not simply extract a rule and re-implement it into SF code.

But Stockfish could build their own neural network implementation for CPU.

--
Srdja