Multiagent Swarms for Korvus (Prelim notes)

Agent Networks

Main value prop is the mqtt (some internet mechanism)

• i think this is how the peer to peer interactions happen
• devise our own method for this for biological interactions
• Begin with an “individual” agent

Three levls - mind, social behaviors, coupling methods

• Mind - profile descriptions (for us it will be biological purpose, chemicals and makeup all of that)

  • profile, memory, reflection, action
  • bio based - profile, past interactions, how it betters or adapts and its main action

• Behaviors - simple and complex

  • Simple - no brainer activities
  • Complex - special case, purposeful
  • Adapt this to our biological agents by having their regular activities and also activites n special case when new condition/new drug introduced.

• Mostly focused on how they can mimick a complex, unpredictable multifaceted human society but biological systems are more deterministic

• each agent with three levels of mental processes - emotion, needs, cognition

• Mobility behavior is the main interaction process

Steps are intention, type, radius and place

• For the bio agents, its gonna be - intended action, where in brain it happens (and its conditions like concentrations, temperature, bs like that), other conditions and the final place
• three types of social relationships (family, friends, colleagues), for us :
  • direct interactors, interactors in that specific pathway, and overall interaction network family

Flows

• Perception flow - agents thoughts and attitudes over time
• Event flow - records events as they happen overtime

Main workflow

• Action determine - same thing
• event feedback - same thing
• memory update - just update interaction history
• emotion/cognition analysis - see how much agent improves based on interaction
• passive/env events - other surrounding events taken notes of

Systems design overview

• LLM API Agents
• MQTT Server
• SQL Db
• MLFlow metrics

Distributed compute

• Group multiple agents in one group and use single proces to execute that group

ABMs

Current Approaches

• Most simulation/optimization algos used in ABM based bio systems are mostly ML based, no LLM application in sight
• Feels like a true untapped idea to merge LLM based networks with ABM Biosystems and be the first person to give a full fledged platform for it

Optimization Methods and Tools

• Incorporate all these algos as agent tools????
• Metaheuristic
  • Particle Swarm, Ant Colony, Bee, Whale, Firefly (different types of optimization algorithms)
• Q-learning from RL - Allows cells to autonomously predict phenotype
• Whole paper is more cell and organelle interaction based since it is oncology focused
• Only 7 out of 104 articles acc used a software framework