A Google DeepMind Technologies research team created a new version of AlphaZero, DeepMind’s foremost chess program, using multiple independent AI systems that trained on different situations. The modified AlphaZero system acted to identify which agent had the best chance of winning and added a “diversity bonus” for pulling strategies from a large selection of choices. When the new system played games, it showed a lot of variety and defeated the original AlphaZero in most matches. The diversified version solved twice as many challenge puzzles and more than half of the total catalog of Penrose puzzles. The idea of creative diversity was praised by Cully for learning to assess and value different approaches. The modified AlphaZero AI had more options for tough situations and more model games to access for better generalization.
Generalizing the findings, a diversified approach can help any AI system and is not limited to reinforcement learning. Diversity was shown to identify promising new drug candidates and develop effective stock-trading strategies. It was suggested that intelligence may be a matter of computational power and creativity could be just a particular computational problem. A diversified approach, while not solving all generalization problems in machine learning, is a step in the right direction. The research resonated with previous efforts showing that cooperation can lead to better performance on hard tasks among humans. Although on the downside, the diverse approach is currently computationally expensive and may not capture the entire spectrum of possibilities.
The whytry.ai article you just read is a brief synopsis; the original article can be found here: Read the Full Article…
