Post AtJ6pO4MClH6ecjC3U by ksaj@infosec.exchange
(DIR) More posts by ksaj@infosec.exchange
(DIR) Post #AtJ6pJtZiuH1i1eWh6 by screwtape@mastodon.sdf.org
2025-04-21T02:38:24Z
0 likes, 0 repeats
@rwxrwxrwx @ksaj was just setting me Braitenberg's VEHICLES to read, and was mentioning using MPI clusters for multiple agent programming. Did you work on anything similar?I haven't got very far in the book yet, but one thing I'm thinking about is how in the first three simple agents in VEHICLES so far, they're all simple sensor-motor systems, which I guess are analogous to how Sandewall uses beliefs-desires-intents - in both cases, it is important that the system operates autonomously.
(DIR) Post #AtJ6pLJUS2b66gmjj6 by rwxrwxrwx@mathstodon.xyz
2025-04-21T03:17:57Z
0 likes, 0 repeats
@screwtape @ksaj I'm unfamiliar with Braitenberg's vehicles but they seem to be close to simplified chemotactic¹ models.It seems that you could easily parallelize their simulation using MPI. Point-to-point communication could be the simplest starting point.¹ https://en.wikipedia.org/wiki/Chemotaxis
(DIR) Post #AtJ6pMXLtzHPtyHJyK by screwtape@mastodon.sdf.org
2025-04-21T03:49:24Z
0 likes, 0 repeats
@ksaj out of interest, Sandewall wrote this agent system immediately after funding dried up for his sparse-communication-with-humans autonomous drones research of the 90s.@rwxrwxrwxI think so. I possibly wrongly associate minds-as-many-but-not-that-many-simple-agents with Minsky, whom I haven't read.Mpi, as I guess all three of us know, is a C library for structuring the syncronizations that have to occur basically, if I can phrase it like that. Though I am using a single shared emacs atm
(DIR) Post #AtJ6pNN6nYgEUUUee0 by screwtape@mastodon.sdf.org
2025-04-21T03:57:53Z
0 likes, 0 repeats
@ksaj @rwxrwxrwx Sandewall's work considers one software-individual to be a cluster of agents sharing a single kernel thread, hence mpi in my Sandewallian software-individuals appears at community level interactions rather than within one individual's agents (which is just where I am right now - actually I'm working on single agents right now).
(DIR) Post #AtJ6pO4MClH6ecjC3U by ksaj@infosec.exchange
2025-04-21T04:59:26Z
0 likes, 1 repeats
@screwtape @rwxrwxrwx That depends on how you use it. MPI is simply message queing and passing. It allows you do do what you've been talking about, but without the shared kernel. Each node is doing it's own thing (which hopefully is synchronized algorithmically), but not always or even necessarily.My experiments were about using flocking algorithms as usually presented, but using mpi to pass messages between the nodes for things that may perturb or be perturbed by the natural flow of the algorithm.For example, one is running out of power. The others now know, and don't rely on it for say, video streaming. It can chill at the tail and do more basic things. Or a need for a specific sensor that only one or some of the nodes/agents/drones have on board, which may require one of them to position themselves at the lead or the tail... etc. There's a lot of things there.One "mission" I had was a swarm that carried explosives on each device, and had various recon/telemetry split amongst them. They would find their way to the target, and go simultaneously kamikaze. Failing the mission, they would return to base, or receive new targeting information. All while running recon to help further incursion. Blowing one or some of them out of the sky would not doom the mission.That's a bit harsh lol, but my first thoughts for this were entirely military. I somehow didn't manage to hear about Sandewall, so I look forward to Tuesday when I'll have some time to look at the resources in the repo.