Re: Concise refutation of halting problem proofs V55 [ halt deciders ]

1 view

olcott

Jan 31, 2022, 12:23:41 PMJan 31
to
On 1/31/2022 11:15 AM, olcott wrote:
> On 1/31/2022 11:05 AM, Mr Flibble wrote:
>> On Fri, 28 Jan 2022 20:34:25 -0500
>> Richard Damon <Ric...@Damon-Family.org> wrote:
>>
>>> On 1/28/22 8:14 PM, olcott wrote:
>>>> Because halt deciders are deciders they are only accountable for
>>>> computing the mapping their actual input finite strings to an
>>>> accept or reject state on the basis of the actual behavior
>>>> specified by these actual inputs.
>>>
>>>
>>> And a Halt Decider, BY DEFINITION, to be correct needs to decide
>>> based on the actual behavior of computaiton the input represents,
>>> which it the equivalent of simulating the input by an ACTUAL UTM
>>> (which H isn't one if it stops simulating before the input reachs a
>>> final state).
>>>
>>>>
>>>> It is like you put a guard on the front door that is supposed to
>>>> report anyone coming in the front door (the actual inputs). Then
>>>> someone comes in the back door (non inputs) and the guard does not
>>>> report this.
>>>
>>> Bad Analogy, the definition of Halting defines what the 'Front Door'
>>> is.
>>>
>>>>
>>>> Since the guard is only supposed to report people coming in the
>>>> front door (actual inputs) it is incorrect to say that the guard
>>>> made a mistake by not reporting people that came in the back door
>>>> (non inputs).
>>>
>>> Right, and if UTM(<H^>,<H^>) halts, then that halting came through
>>> the front door unless you are lying about working on the Halting
>>> Problem.
>>>
>>>>
>>>> Ĥ.q0 ⟨Ĥ⟩ ⊢* Ĥ.qx ⟨Ĥ⟩ ⟨Ĥ⟩ ⊢* Ĥ.qy ∞
>>>> Ĥ.q0 ⟨Ĥ⟩ ⊢* Ĥ.qx ⟨Ĥ⟩ ⟨Ĥ⟩ ⊢* Ĥ.qn
>>>>
>>>> The copy of Linz H at Ĥ.qx (embedded_H) determines the halt status
>>>> of its input on the basis of whether or not the pure simulation of
>>>> any finite number of steps of this input can possibly ever reach a
>>>> final state of this simulated input.
>>>
>>> And you have yet to prove that this is ACTUALLY possible. In fact,
>>> this "ANSWER' is precisely the fallacy of assuming the conclusion.
>>> You are basically saying that you can make a Halt Decider, because if
>>> you have a Halt Decider you can decide on Halting.
>>>
>>>>
>>>> When embedded_H correctly determines that its simulated input would
>>>> never reach its final state it aborts its input and transitions to
>>>> Ĥ.qn.
>>>
>>> Again, you are assuming something you have not proved, and has been
>>> proved to be impossible in this case. This is more of your Fairy Dust
>>> Powered Unicorns.
>>>
>>> FAIL.
>>>
>>>> When this causes Ĥ applied to ⟨Ĥ⟩ to halt that makes no difference
>>>> because the guard is only accountable for watching the front door.
>>>
>>> Except that if H^ halts because the copy of H aborts it simulaton of
>>> a copy of H^, then this halting IS the 'Front Door' that the guards
>>> were responsible to detect.
>>>
>>> Apparently they were asleep on the trying to make up a story to cover.
>>>
>>> FAIL.
>>>
>>>
>>>> https://www.researchgate.net/publication/358009319_Halting_problem_undecidability_and_infinitely_nested_simulation_V3
>>>>
>>>>
>>>>
>>>
>>> We have gone over this many times, it is clear that you are just
>>> lying that you are working on the Halting Problem because you refuse
>>> to use the actual definitions of Halting from the problem, but try to
>>> shade it with weasle words to allow you to try to sneak in a false
>>> premise.
>>>
>>> Either that or you are just too mentally deficient to be capable of
>>> doing any real logic, and likely should be committed to keep yourself
>>> from being a danger to yourself.
>>>
>>> FAIL.
>>
>> You both need to be sectioned IMO. Give it a fucking rest.
>>
>> /Flibble
>>
>
> This is my lifetime legacy and the FLIPI index projects that I will die
> by next December.
>
> https://www.mdcalc.com/follicular-lymphoma-international-prognostic-index-flipi
>
>
> Halting problem undecidability and infinitely nested simulation (V3)
>
> https://www.researchgate.net/publication/358009319_Halting_problem_undecidability_and_infinitely_nested_simulation_V3

Once it is understood that I am correct this opens up a whole new world
for AI research:

(A) Computation will be understood to have truly unlimited potential.

(B) Davidson's truth conditional semantics will finally be anchored in a
correct formal definition of truth, refuting the Tarski Undefinability
theorem.

--