ROCCA Weekly Community Call (ROCCA-WCC)

[Nil Geisweiller]

Hi everyone,

I should have soon more time for ROCCA

https://github.com/opencog/rocca
https://github.com/singnet/rocca (mirror)

I plan on hosting a community call every Friday for anyone interested,
but the time remains to be determined.

My availability would be

8am-3pm (EEST) (ideal, cause I'd be systematically available)
4pm-10pm (EEST) (less ideal but possible)

Any preference please let me know.

Nil

[Nil Geisweiller]

On 6/22/21 6:50 PM, Nil Geisweiller wrote:
> My availability would be
>
> 8am-3pm (EEST) (ideal, cause I'd be systematically available)
> 4pm-10pm (EEST) (less ideal but possible)

Actually it'd be more like

8am-3pm (EEST) (ideal, cause I'd be systematically available)

7pm-10pm (EEST) (less ideal but possible)

Nil

[Adrian Borucki]

My availability would be from 11:00 CEST (so 12:00 EEST) to 19:00 CEST (so 20:00 EEST).

[Nil Geisweiller]

Thanks Adrian,

I'm thinking of making it monthly for now.

I suggest 3pm EEST, every 1st Friday of the month. Starting either
August or September, depending on my summer vacations. I'll send more
info prior to the call.

Nil

> --
> You received this message because you are subscribed to the Google
> Groups "opencog" group.
> To unsubscribe from this group and stop receiving emails from it, send
> an email to opencog+u...@googlegroups.com
> <mailto:opencog+u...@googlegroups.com>.
> To view this discussion on the web visit
> https://groups.google.com/d/msgid/opencog/9f2af87c-6de9-4f55-8ca3-483ab99a4de3n%40googlegroups.com
> <https://groups.google.com/d/msgid/opencog/9f2af87c-6de9-4f55-8ca3-483ab99a4de3n%40googlegroups.com?utm_medium=email&utm_source=footer>.

[Patrick Hammer]

Hi Nil!

ROCCA looks very interesting!

If you don't mind I will also join the meeting (I have more time now as I just defended my dissertation about ONA), but as passive observer for now when you don't mind, as I'm not very familiar yet with OpenCog and what you did.

Is there a description of what cognitive functions the ROCCA agent does already possess? Or a related publication?

Especially, I'm interested in which representations it is already able to learn at runtime, such as new behaviors corresponding to the "famous" triple (antecedent action) => consequent.

In Cartpole this kind of behavior seems to be given via PredictiveImplicationScopeLink: https://github.com/opencog/rocca/blob/c9f24ca76ab710eeff939359d84bd90b7acb11ab/examples/cartpole.py#L172

But in the other examples it's already learning such hypotheses by itself?

Best regards,

Patrick

[Adrian Borucki]

Hi Patrick,

I can answer some of your questions to some extent:

- The basic Cartpole example uses hardcoded schemas (we call these context & action => goal triples cognitive schemas) and indeed the other examples try to learn useful ones by themselves.

  - Currently most of the heavy lifting is done by the Pattern Miner. Basically, it will produce some LambdaLinks based on some surprising patterns that it can spot from the database of observations and actions taken by the agent, that then get transformed into PredictiveImplicationScopeLinks that constitute cognitive schemas.

- You can find two relevant papers in the code comments about the use of Thompson Sampling (http://auai.org/uai2016/proceedings/papers/20.pdf) and Partial Operator Induction (this one written by Nil, here).

As for me, so far I haven’t been able to get it to achieve some tangible goals, like solving the Cartpole-v1 environment. For Minecraft, one of the easier tasks is the navigation challenge which requires the agent to reach and touch some block. Not much progress in this regard so far as well.

There is also a document outlining current state of the project and potential roadmap written by Nil here.

Best regards,

Adrian

[Nil Geisweiller]

Hi guys,

thanks for your reply, Adrian. To add to that, there's a terribly drafty
document about temporal reasoning in the pln repo

https://github.com/opencog/pln/blob/master/opencog/pln/rules/temporal/temporal-reasoning.md

It'd be a delight to have you join the call, Patrick. I'll send a
reminder with the place and time (should still be Fri 6 Aug 3pm EEST, if
that works for everybody) in a separate email.

Nil

On 7/25/21 4:12 PM, Adrian Borucki wrote:
> Hi Patrick,
>
> I can answer some of your questions to some extent:
> - The basic Cartpole example uses hardcoded schemas (we call these

> /context & action => goal/ triples /cognitive schemas/) and indeed the

> other examples try to learn useful ones by themselves.
>   - Currently most of the heavy lifting is done by the Pattern Miner.

> Basically, it will produce some /LambdaLink/s based on some surprising

> patterns that it can spot from the database of observations and actions
> taken by the agent, that then get transformed into

> /PredictiveImplicationScopeLink/s that constitute cognitive schemas.

> - You can find two relevant papers in the code comments about the use of
> Thompson Sampling (http://auai.org/uai2016/proceedings/papers/20.pdf)
> and Partial Operator Induction (this one written by Nil, here

> <https://github.com/ngeiswei/papers/blob/master/PartialBetaOperatorInduction/PartialBetaOperatorInduction.pdf>).

>
> As for me, so far I haven’t been able to get it to achieve some tangible
> goals, like solving the Cartpole-v1 environment. For Minecraft, one of
> the easier tasks is the navigation challenge which requires the agent to
> reach and touch some block. Not much progress in this regard so far as well.
>
> There is also a document outlining current state of the project and
> potential roadmap written by Nil here

> <https://github.com/opencog/rocca/blob/master/doc/proto-agi-early-progress-report-and-planning.md>.

>
> Best regards,
> Adrian
>
> On Sunday, 25 July 2021 at 05:56:01 UTC+2 pat...@gmail.com wrote:
>
> Hi Nil!
>
> ROCCA looks very interesting!
> If you don't mind I will also join the meeting (I have more time now
> as I just defended my dissertation about ONA

> <https://github.com/opennars/OpenNARS-for-Applications>), but as

> passive observer for now when you don't mind, as I'm not very
> familiar yet with OpenCog and what you did.
> Is there a description of what cognitive functions the ROCCA agent
> does already possess? Or a related publication?
> Especially, I'm interested in which representations it is already
> able to learn at runtime, such as new behaviors corresponding to the

> "famous" triple *(antecedent action) => consequent*.

>
> In Cartpole this kind of behavior seems to be given via

> *PredictiveImplicationScopeLink*:
> https://github.com/opencog/rocca/blob/c9f24ca76ab710eeff939359d84bd90b7acb11ab/examples/cartpole.py#L172

> <https://groups.google.com/d/msgid/opencog/9f2af87c-6de9-4f55-8ca3-483ab99a4de3n%40googlegroups.com?utm_medium=email&utm_source=footer
> <https://groups.google.com/d/msgid/opencog/9f2af87c-6de9-4f55-8ca3-483ab99a4de3n%40googlegroups.com?utm_medium=email&utm_source=footer>>.

>
>
> --
> You received this message because you are subscribed to the Google
> Groups "opencog" group.
> To unsubscribe from this group and stop receiving emails from it, send
> an email to opencog+u...@googlegroups.com
> <mailto:opencog+u...@googlegroups.com>.
> To view this discussion on the web visit

> https://groups.google.com/d/msgid/opencog/f138aca8-3416-48ba-af20-3e1e9a35770en%40googlegroups.com
> <https://groups.google.com/d/msgid/opencog/f138aca8-3416-48ba-af20-3e1e9a35770en%40googlegroups.com?utm_medium=email&utm_source=footer>.

[Patrick Hammer]

Hi Nil and Adrian!

Thank you for your detailed responses and clarifications! I have finally read through both papers Adrian suggested.

Thompson Sampling makes perfect sense, we are using similar principles in the "NAL world" to balance exploration&exploitation according to how promising (certainty to lead to the desired outcome) the available options are.

And partial operator induction looks very reasonable as well since it's way more practical to evaluate certainties considering only the options which contextually apply, even though it doesn't give an exact probability value or efficient-to-calculate error bound even. Drescher suggested a similar, but simpler, practical decision for his Schema Mechanism (1986): to simply assume true/false conditions and to keep track of success and failures when schemas have their context fulfilled (the antecedent condition is true), he named the resulting success/(success+failures) measure "reliance", which he directly used to chose among competing options in the chaining / planning process. Of course his simplistic true/false treatment can be an issue if observations themselves are uncertain. I wouldn't worry about this for now though, since PLN & the usage of beta distributions will likely allow to address this in a better way.

The ideas and inference rules presented in the temporal reasoning document make sense to me as well, hence I'm also optimistic that you will soon be able to make it all work!

Thanks again, and I'm looking forward to meeting you on August 6!

Best regards,

Patrick

[Nil Geisweiller]

Oh, I didn't know about Gary Drescher's Schema Mechanism (thanks for the
ref, Patrick!) which is very much like Ben's Cognitive Schematic. Like
Gary, Ben mentions Jean Piaget a lot, so I guess they both got that from
him (and I suppose they're both aware of their respective wordings and
I'm just late to the party).

Nil

> <http://auai.org/uai2016/proceedings/papers/20.pdf>)

> <https://groups.google.com/d/msgid/opencog/f138aca8-3416-48ba-af20-3e1e9a35770en%40googlegroups.com?utm_medium=email&utm_source=footer
> <https://groups.google.com/d/msgid/opencog/f138aca8-3416-48ba-af20-3e1e9a35770en%40googlegroups.com?utm_medium=email&utm_source=footer>>.

>
>
> --
> You received this message because you are subscribed to the Google
> Groups "opencog" group.
> To unsubscribe from this group and stop receiving emails from it, send
> an email to opencog+u...@googlegroups.com
> <mailto:opencog+u...@googlegroups.com>.
> To view this discussion on the web visit

> https://groups.google.com/d/msgid/opencog/98b4b45f-9dad-4d1a-abdb-b350c13f622cn%40googlegroups.com
> <https://groups.google.com/d/msgid/opencog/98b4b45f-9dad-4d1a-abdb-b350c13f622cn%40googlegroups.com?utm_medium=email&utm_source=footer>.