Reinforcement Learning
- Explore:
- Overview
- Our Experience
Reinforcement learning uses trial and error feedback to teach machines how to achieve a goal. Metron develops AI systems that teach themselves.
Join Our Team
Metron is a leader in researching and developing techniques to scale reinforcement learning algorithms to large state and action spaces enabling 21st century strategic and tactical mission planning and battle management.
Action and Reward
Reinforcement learning uses trial and error feedback to teach machines how to achieve a goal. Applications include self-driving cars, playing video games, robotic manipulation, and strategy games. The simplest form of reinforcement learning involves an agent taking an action and observing a reward or penalty based on that action. For example, choosing a restaurant is an action, and the rating based on the experience is the reward. That rating (along with previous scores) affects how likely the agent will choose that restaurant in the future.
More sophisticated reinforcement learning problems can be addressed with Markov decision process models, and solved using stochastic, dynamic programming methods or Q-learning. Metron scientists specialize in selecting and applying the best mathematical methods to solve our clients’ toughest problems.
Our Experience
We teach AI systems to teach themselves.
Metron has successfully used stochastic, dynamic programming and Q-learning approaches to learn effective policies for a variety of applications. Part of this research includes how to scale these algorithms to large state and action spaces. We are also investigating multi-agent learning problems in which multiple players must learn and interact within the same environment. The resulting joint spaces are too large for traditional methods, so new techniques are required.
Metron is researching and developing reinforcement learning advancements that will allow the application of these techniques to large scale strategic and tactical mission planning and battle management. This work is enabled by our rich campaign simulation and modeling capabilities in Cyber Assassin and NSS.
Related Content
Bayesian Optimization
Bayesian optimization is a cost-effective method for solving "black box" function-optimization problems in machine learning by strategically choosing a highly informative set of points to sample.
Explore Bayesian Optimization
Metaheuristics
Metron applies advances in artificial intelligence to reduce the computational cost required for searching complex solution spaces.
Explore Metaheuristics
Nonlinear Gradient Descent
Metron scientists use nonlinear gradient descent methods to find optimal solutions to complex resource allocation problems and train neural networks.
Explore Nonlinear Gradient Descent
Stochastic Dynamic Programming
Metron data scientists apply stochastic dynamic programming methods to solve incredibly large and complex planning problems.
Explore Stochastic Dynamic Programming
System of Systems
Metron is a pioneer in advanced decision aid technologies for composing systems of systems and mission plans.
Explore System of Systems
Metron Careers
Decision Support Career Opportunities
Metron hires research scientists with experience developing novel approaches that advance the state of the art in mathematics and artificial intelligence. Our scientists work alongside subject matter experts applying these innovations to new problem domains.