A Multi-Agent Reinforcement Learning Architecture for MQL5 Trading

Summary

The article describes an MQL5 trading system organized around memory units, individual agents called brains, and a collective decision layer. Each memory unit associates an encoded market state with an action and tracks reward, estimated value, activations, success rate, and importance. Agents can have slightly different learning and exploration settings; their decisions are combined with weights tied to prior performance. The system represents actions on a continuum between selling and buying, and uses a feature set of normalized indicators and price changes to characterize market conditions.

The design also includes mechanisms for balancing buy and sell behavior, pruning or managing stored memories, and saving or reloading learned models. The article explains these components as a way to support continued learning and varied agent behavior, but the supplied text gives no quantified trading results or comparative evaluation to show that the architecture is profitable or robust. It identifies computation demands, growth in memory size, overfitting, and changing market conditions as limitations. Its descriptions of adaptation and diversification should therefore be read as design goals rather than demonstrated performance.

Key ideas

• Market states are encoded into memory units that associate conditions with actions and update reward-related metrics.
• Multiple agents with varied learning settings contribute to a collective trading decision.
• The described feature set combines normalized technical indicators with price changes across different windows.
• The design includes model persistence and memory management, while adding agents and memories can increase computational load.
• The supplied material offers no quantified performance evidence and identifies overfitting and regime changes as key risks.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.