Evolutionary Reinforcement Learning is an exciting development in machine learning that promises significant advances. It combines the strengths of two distinct approaches: reinforcement learning and evolutionary computation. By implementing trial-and-error learning from reinforcement learning and natural selection from evolutionary algorithms, it creates an effective tool for artificial intelligence development. A review paper was recently published in Intelligent Computing and it shed light on the progress and the most advanced methods for integrating evolutionary computation with reinforcement learning.
Reinforcement learning has had great successes so far; for example AlphaGo and, more recently, Google DeepMind’s soccer playing robots. Despite that, it still has issues with exploration and exploitation trade-off, reward design, generalization, and credit assignment. That’s why evolutionary computation, which draws from natural selection’s process of solving problems, is an attractive solution to some reinforcement learning problems.
Evolutionary reinforcement learning contains six main research areas:
• Evolution Strategies
• Sparse Reward Learning
• Multi-Objective Reinforcement Learning
• Genetically Augmented Neural Networks
• Co-Evolution
• Meta-Learning.
Evolutionary reinforcement learning can solve complex reinforcement learning tasks, even with rare or misleading rewards. It carries the downside of requiring heavy computation power – something that researchers are actively trying to improve in terms of efficiency. This includes encoding, sampling, search operators, algorithmic frameworks, and evaluation.
As evolutionary reinforcement learning is still an ongoing field of research, there are exciting development prospects. Improvements to computational efficiency and the exploration of new benchmarks, platforms, and applications can make the process even more effective and useful. With the potential to break through boundaries in the reinforcement learning field, evolutionary reinforcement learning offers immense opportunities for artificial intelligence development.
