Smart active particles learn and transcend bacterial foraging strategies

Area of Science:

• Microbial Ecology
• Artificial Intelligence
• Biophysics

Background:

• Bacteria and microorganisms evolve efficient foraging strategies in unknown environments.
• Statistical models describe microbial movement, but the optimality of learned strategies remains unclear.
• A lack of methods to develop and compare alternative strategies hinders understanding.

Purpose of the Study:

• To use deep reinforcement learning (DRL) to develop and analyze bacterial foraging strategies.
• To compare DRL-learned strategies with known bacterial chemotaxis models.
• To assess the survival and foraging advantages conferred by learned strategies.

Main Methods:

• Employing deep reinforcement learning (DRL) to simulate a run-and-tumble agent.
• Training the agent to optimize nutrient foraging and survival.
• Analyzing the emergent motion patterns and tumble rate distributions.

Main Results:

• The DRL agent learned motion patterns highly similar to chemotactic bacteria.
• Significant differences were observed in the learned tumble rate distribution compared to standard models.
• These learned differences provided the agent with superior foraging and survival capabilities.

Conclusions:

• DRL offers a powerful method for discovering efficient search strategies in unknown environments.
• Learned strategies can outperform traditional models, highlighting adaptive advantages.
• This approach has potential applications in programming microswimmers, nanorobots, and active particles for various tasks.