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 On Wednesday, July 28, 2021 at 11:32:03 AM UTC Nil wrote: > 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 [email protected] 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://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 > > > > On Monday, July 12, 2021 at 1:01:24 PM UTC Nil wrote: > > > > 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 > > > > On 6/24/21 11:10 PM, Adrian Borucki wrote: > > > My availability would be from 11:00 CEST (so 12:00 EEST) to > > 19:00 CEST > > > (so 20:00 EEST). > > > > > > On Tuesday, 22 June 2021 at 23:22:12 UTC+2 Nil wrote: > > > > > > 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 > > > > > > -- > > > 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 [email protected] > > > <mailto:[email protected]>. > > > 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 > > > > > > > > > < > 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 [email protected] > > <mailto:[email protected]>. > > 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 > >. > -- 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 [email protected]. To view this discussion on the web visit https://groups.google.com/d/msgid/opencog/98b4b45f-9dad-4d1a-abdb-b350c13f622cn%40googlegroups.com.
