3 ms·
A primary problem in marrying logic to ML is doing so in a differentiable way. i.e. when your SAT part of the system is suboptimal, how do you know which parame
by 6gvONxR4sf7o 7y ago
A primary problem in marrying logic to ML is doing so in a differentiable way. i.e. when your SAT part of the system is suboptimal, how do you know which parameters of the learned part to change, and in which direction?
- nl 7y agoWe propose Differentiable Satisfiability and Differentiable Answer Set Programming (Differentiable SAT/ASP) for multi-model optimization. Models (answer sets or satisfying truth assignments) are sampled using a novel SAT/ASP solving approach which uses a gradient descent-based branching mechanism. https://arxiv.org/abs/1812.11948 https://arxiv.org/abs/1812.11948
- smallnamespace 7y agoFrom the article, a lot of the heuristics seems to revolve around and 1) which variables to try first 2) which learned clauses to keep in your set. Assign scores and differentiate on those scores. This is (very loosely) how AlphaGo also works - you take a tree-based algorithm (MCTS) and then use neural nets to score the next node to expand.
- mlevental 7y agoyou don't need differentiability you just need a learning rule. there are learning algorithms that don't use gradient ascent.