Global convergence of policy gradient methods to (almost) locally optimal policies

Mathematics

• Optimal Policy 95%
• Gradient Method 92%
• Policy 80%
• Nonconvex Optimization 39%
• Reward 28%
• Reinforcement Learning 24%
• Sweden 12%
• Gradient 12%
• Infinite-horizon Problems 11%
• Stochastic Gradient 10%
• Infinite Horizon 9%
• Saddle 8%
• Strictly positive 8%
• Horizon 8%
• Alternatives 5%

Engineering & Materials Science

• Gradient methods 100%
• Reinforcement learning 27%
• Parameterization 14%
• Machine learning 10%