Consensus Control of Micro Multi-Agent Reinforcement Learning Systems for Tumor Treatment

A consensus control algorithm for micro multi-agent reinforcement learning (MMARL) systems, including a collision avoidance system (CAS) is presented in this paper to guarantee the safety of surrounding healthy tissues in the human body. Our proposed method employs multiple micro-agents to enhance treatment accuracy and efficiency in targeting and destroying cancer tumors within a complicated biological environment. Designing an MMARL system, developing a consensus control strategy considering CAS, and validating it using simulations focusing on potential in vivo applications have been presented here. The simulation results indicate that agent <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$A_{1}$</tex> showed the most consistent improvement, with 1.26% rewards increasing, while agent <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$A_{2}$</tex> displayed more variability, with 1.84% rewards raising. Moreover, agent <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$A_3$</tex> stabilized with 1.25% rewards boosting. Additionally, the agents displayed similar speed patterns, with agent <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$A_1$</tex> having the lowest variance in speed (∼ 0.1) and agent <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$A_{3}$</tex> the highest (∼ 0.2), reflecting differences in stability. The desired results contain increased targeting accuracy and minimized side effects, providing a non-invasive treatment rather than surgical procedures for high-risk patients.