Infinity/NaN output

[dan-ryan]

An old Neataptic bug which is still bugging me.

I'm getting Infinity/NaN outputs while training. Using the latest Node.js.

My settings:

    Methods.mutation.MOD_ACTIVATION.mutateOutput = false;
    Methods.mutation.MOD_ACTIVATION.allowed = [
        Methods.activation.LOGISTIC,
        Methods.activation.TANH,
        Methods.activation.STEP,
        Methods.activation.SOFTSIGN,
        Methods.activation.SINUSOID,
        Methods.activation.GAUSSIAN,
        Methods.activation.BIPOLAR,
        Methods.activation.BIPOLAR_SIGMOID,
        Methods.activation.HARD_TANH,
        Methods.activation.INVERSE,
        Methods.activation.SELU,
        Methods.activation.RELU,
        Methods.activation.BENT_IDENTITY,
        Methods.activation.IDENTITY,
    ];

    neat = new Neat(4, 1, null,
        {
            mutation: Methods.mutation.ALL,
            popsize: 100,
            mutationRate: 0.2,
            elitism: 10,
            equal: true,
            network: new Architect.Random(4, 5, 1),
            provenance: 2,
            maxNodes: 7, 
            maxConns: 10,
            maxGates: 5,
        }
    );

Infinity Example:

The normalised data:
[0.9354838709677419, 0.5, 0.5933786078098472, 0.5880669161907702]

The genome:

{"nodes":[{"bias":0,"type":"input","squash":"LOGISTIC","mask":1,"index":0},{"bias":0,"type":"input","squash":"LOGISTIC","mask":1,"index":1},{"bias":0,"type":"input","squash":"LOGISTIC","mask":1,"index":2},{"bias":0,"type":"input","squash":"LOGISTIC","mask":1,"index":3},{"bias":1.1508084667830487,"type":"output","squash":"SELU","mask":1,"index":4}],"connections":[{"weight":1,"from":4,"to":4,"gater":4},{"weight":-0.05001360658035439,"from":3,"to":4,"gater":null},{"weight":0.9984137443904727,"from":2,"to":4,"gater":null},{"weight":-0.7832753538521565,"from":1,"to":4,"gater":null},{"weight":-0.9040067054346645,"from":0,"to":4,"gater":4}],"input":4,"output":1,"dropout":0}

NaN Example:

The normalised data:
[0.3870967741935484, 0.75, 0.5040295048190726, 0.5575469079452833]

The genome:

{"nodes":[{"bias":0,"type":"input","squash":"LOGISTIC","mask":1,"index":0},{"bias":0,"type":"input","squash":"LOGISTIC","mask":1,"index":1},{"bias":0,"type":"input","squash":"LOGISTIC","mask":1,"index":2},{"bias":0,"type":"input","squash":"LOGISTIC","mask":1,"index":3},{"bias":-0.0717463180924848,"type":"hidden","squash":"BENT_IDENTITY","mask":1,"index":4},{"bias":-0.02994106373052867,"type":"output","squash":"LOGISTIC","mask":1,"index":5}],"connections":[{"weight":1,"from":4,"to":4,"gater":null},{"weight":1,"from":5,"to":5,"gater":null},{"weight":-0.048341462482942354,"from":4,"to":5,"gater":null},{"weight":-0.4890970041600281,"from":3,"to":5,"gater":5},{"weight":0.9984137443904727,"from":2,"to":5,"gater":5},{"weight":-0.4890970041600281,"from":3,"to":4,"gater":4},{"weight":0.024574079733371557,"from":1,"to":5,"gater":null},{"weight":0.9984137443904727,"from":2,"to":4,"gater":4},{"weight":-0.9040067054346645,"from":0,"to":5,"gater":5},{"weight":0.049181674792330154,"from":1,"to":4,"gater":null},{"weight":0.04742932010695605,"from":0,"to":4,"gater":null}],"input":4,"output":1,"dropout":0}

wagenaartje/neataptic#130

[luiscarbonell]

@dan-ryan It looks like we're carrying a few of these over.

@christianechevarria and I are a little caught up in some regular life things till Wednesday, but I'll try to squeeze out as much time as can for this one. I'm also not a fan of the NaN/Infinity bugs.

[dan-ryan]

There was an attempt in the past to fix NaNs.
wagenaartje/neataptic@2fd4dbd

[luiscarbonell]

@dan-ryan I'm having a hard time replicating the Infinity/NaN bug.

The genomes that you placed, where they for the Neat.getAverage() or Neat.getFittest()?

[dan-ryan]

@luiscarbonell
What I could do is share my project class that way you can debug this. What's your email?

[luiscarbonell]

@luiscarbonell
What I could do is share my project class that way you can debug this. What's your email?

[email protected]

[dan-ryan]

@luiscarbonell I've sent you an email. A simple enough example to get running and able to get NaNs and Infinities quickly.

[christianechevarria]

Update:

[luiscarbonell]

This issue is really a mutation and node.squash issue.

We need to be able to mutate activation functions on output neurons - without creating new ones.

This is sticky-water because of ranges (-1, 1) or (0,1) or (-Infinity, Infinity)

[christianechevarria]

@dan-ryan While playing around with a logic gate demo for the repo link ran into a silent bug that could have a hand in this, turns out that architect.LSTM is spitting out NaNs, Random is failing too.

This could be related to #35 as a common root cause

Seems like a suite of tests for architect and a potential re-write is in order

[christianechevarria]

Added a clamp to activation functions in 71301ba that keeps Infinity to Number.MAX_VALUE and -Infinity to -Number.MAX_VALUE so that they won't propagate through the network.

Although this should solve the Infinity / NaNs issue, it will create some loss of learning ability in cases where activations really should exceed the built-in numerical limits of JS.

The ideal solution would likely involve finding / implementing a big number library that would allow us to operate on numbers of arbitrary length