Concurrent and Distributed NARS

[djire...@gmail.com]

Hi everybody,

I've been experimenting with NARS running on the BEAM virtual machine and using the Erlang's Open Telecom Platform (OTP). 

This  gives concurrency and distributed computing (along with fault tolerance) out-of-the-box (of course not without it's own problems). 

I was thinking about writing it up for the next AGI Conf 2026, but wondered if there were any other attempts at a concurrent / distributed NARS I could refer to. I recall there were at least 50 implementations (from a slide presented by someone else) but couldn't find any distributed implementations nor were they mentions in this google group. 

I was considering to heavily cite  "The Event Buffer of OpenNARS" paper which had a parallel architecture to it. 

Appreciate any thoughts. 

Regards,

David

 

 

[Pei Wang]

Hi David,

The relevant works I can think of at the moment:

• Intelligent Multiagent Systems based on Distributed Non-Axiomatic Reasoning
• https://github.com/rasom/opennars2

There have been discussions and experiments on multi-agent cooperation and hardware acceleration for OpenNARS and ONA; however, I have not found any publicly available writing or code.

Looking forward to your results!

Regards,

Pei

--
You received this message because you are subscribed to the Google Groups "open-nars" group.
To unsubscribe from this group and stop receiving emails from it, send an email to open-nars+...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/open-nars/9eca5628-0bc4-42bf-a274-c64a4a057f0fn%40googlegroups.com.

[David Ireland]

Thanks Pei.  I hadn't seen Dejan Mitrovic's thesis. 

To view this discussion visit https://groups.google.com/d/msgid/open-nars/CAK0i%3Dhhb0LNYhQ%2B-kmeZBk-u8KsXpeYnO-LuDbDEoz8xDMa%2BXQ%40mail.gmail.com.

[Patrick Hammer]

Hi David!

The most advanced distributed NARS implementation so far is Tony Lofthouse's ALANN2018, which scales NARS (NAL-1 to NAL-6) to thousands of threads operating concurrently on shared concept memory with high efficiency: https://github.com/opennars/ALANN2018

However, temporal and procedural reasoning are much harder to handle in a distributed architecture, and ALANN2018 did not address them. More generally, these aspects have also long been the weakest parts of NARS conceptually, not just implementation-wise, with many missing and underspecified ideas, making embodied NARS-based divert far from AGI expectations.

I recently tried to push beyond ONA's sensorimotor capabilities, with some success, focusing on improved temporal and event-stream reasoning but with NAL1-8 support: https://github.com/patham9/estream

While limited, NAL-based sensorimotor can still be useful though, as it can allow agents to adapt very effectively when event encodings are at the right abstraction level.

Best regards,

Patrick

[David Ireland]

Hi Patrick,

Thanks! I've been meaning to read Tony's book. 

Am I right in assuming the temporal issues in a distributed cluster are because it would be incredibly difficult to synchronise a time stamp? 

Are you also able to elaborate on some of the other weaker aspects?

Thanks.

Regards,

David

 

--

You received this message because you are subscribed to the Google Groups "open-nars" group.
To unsubscribe from this group and stop receiving emails from it, send an email to open-nars+...@googlegroups.com.

To view this discussion visit https://groups.google.com/d/msgid/open-nars/CA%2BK1R6WY7N%3Dma6G4xxO3JaQJKsRUmWqUrnkYAcCYzgWfFVdm-Q%40mail.gmail.com.

[Patrick Hammer]

Hi David!

I just saw that you responded, sorry for the delay.

This is mainly a coordination problem. To detect temporal patterns, large numbers of events must be combined during reasoning. In a distributed design where events are routed to separate concepts, this quickly breaks down: coordination becomes difficult, and some form of centralized structure (buffers, FIFOs, etc.) becomes necessary.

Independent of AIKR, the scale of event integration truly matters. At high well-prepared/engineered abstraction levels, relations like <a =/> b> with single events can be useful. But for high-throughput, sensory-like streams, even sequences or parallel conjunctions of ~20 events are often insufficient. Dependencies between events are often way more complex than simply "sequence of events A predicts B", and behavior demands extremely fine-grained, tightly coordinated action, something that is undermined both by distributed system realities and by the constraints of logical descriptions.

Best regards,

Patrick