AI Technology: Machine Learning; Natural Language Processing; Deep Learning; Vision (image recognition, image generation, text recognition OCR, etc.); Audio (speech recognition, speech synthesis, etc.); Virtual agents; Decision management: Classified and statistical learning methods.
This project trains a foundational AI system that enables autonomous agents to trade with each other based on the forecasted value of digital interactions like attention, service events, or decentralized asset flows.
These interactions, while diverse in form, share a common economic structure: their value unfolds probabilistically over time, shaped by behavioral context and governed by uncertain future outcomes (e.g., conversion, payment, or engagement).
The objective is to develop an AI architecture that learns this abstract structure of yield-bearing workflows and simulates dynamic pricing strategies through multi-agent reinforcement learning.
The study aims to enable:
- Coordinated AI agent behavior across engagement and trading scenarios
- Attention and engagement-based pricing optimization
- AI-driven valuation of behavioral and decentralized finance (DeFi) interactions
- Adaptive recommendation engines for live shopping, fan engagement, and real-time commerce.
While many current AI systems are designed to optimize short-term engagement or monetization, this project takes a fundamentally different approach. We aim to build a structured, cooperative AI marketplace governed by alignment protocols where agents dynamically price and trade based on forecasted economic value.
This represents a shift from isolated model optimization to simulating real-time, stakeholder-aligned economic coordination between AI agents.
Our training pipelines will be optimized for distributed execution, memory-efficient long-sequence processing, and scalable multi-agent learning.
By building this stack natively for ROCm, the project contributes directly to EuroHPC’s mission of sovereign, open, and scalable AI infrastructure.
Transformer-based deep learning models such as BERT-style architectures implemented via Hugging Face Transformers and optimized through Optimum-AMD are used to generate probabilistic forecasts of the value of workflow-bound events, including attention spikes, asset issuances, and service completions.
These models are trained on high-frequency event streams using long-sequence encoding, contextual signal integration, and temporal abstraction layers that capture risk-adjusted behavioral patterns over time.
Reinforcement Learning (RL), including Multi-Agent RL (MARL) frameworks, is applied to simulate dynamic valuation and resource allocation strategies. Agents learn to adapt their pricing policies in response to decaying value signals, demand fluctuations, and real-time engagement feedback.
The system supports both cooperative and competitive agent configurations to model a wide range of economic scenarios.Training pipelines are implemented using PyTorch (ROCm-compatible release) and optimized for LUMI-G’s MI250X architecture.
The stack supports model and data parallelism, mixed precision, and memory-efficient attention mechanisms (e.g., Triton or DeepSpeed ROCm integration), enabling efficient long-sequence transformer training across 16–40 GPUs. Explainability and interpretability are implemented using ROCm-compatible tools such as SHAP, Captum, and custom attention attribution techniques.
While these add development overhead regardless of infrastructure, they reflect our commitment to a sovereign, auditable AI pipeline.
This approach ensures that decisions made by forecasting models and pricing agents are transparent, reproducible, and aligned with EU standards for fairness and accountability, reinforcing long-term trust in AI-driven economic reasoning.
Åsa Sundqvist, Caprendum AB, Sweden