IAS Seminar "From AlphaChem to AlphaAnything: Learning To Plan Chemical Synthesis and Beyond"

Speaker:

Dr. Mike Preuss, Fakultät Wirtschaftsinformatik und Statistik, Universität Münster

Abstract:

Monte Carlo Tree Search (MCTS) and Deep Neural Networks (DNN) in combination have pushed the limits of what artificial intelligence (AI) can do in areas where humans have been perceived as dominant over machines, as for the game Go. However, so far this has been limited to domains with simple and known rules, such as board games, where perfect world knowledge and cheap environment simulators are available.

Unfortunately, in contrast to environments provided by board games, in the most important real-life domains, it is either nontrivial to comprehensively write down the rules, or the rules are simply yet unknown. To provide data from interaction with the environment, it is therefore usually required to perform very costly experiments or execute complex, albeit still unprecise simulations.

Chemical retrosynthesis (you know the product, but not how to get there) is one of such real-life domains with highly non-trivial, partially unknown rules. We show that this problem can very effectively be tackled with MCTS constrained by DNNs that learn from essentially the complete history of organic chemistry, as described in a recent Nature paper [1]. Our algorithm produces plans which human chemists could not tell apart from established plans taken from the literature. But it does not have to end here. We attempt to generalize the approach in order to make it applicable to other research areas as well.

[1] Segler, M. H., Preuss, M., & Waller, M. P. (2018). Planning chemical syntheses with deep neural networks and symbolic AI. Nature, 555(7698), 604. https://www.nature.com/articles/nature25978

Anyone interested is cordially invited to participate in this seminar.

Contact: Dr. Jenia Jitsev, JSC