Transforming logistics with self-learning AI

Source: NVIDIA blog. NVIDIA Inception programme member InstaDeep is using the same techniques that made AlphaZero successful to make logistics more efficient.

One of the longest-running challenges in the logistics industry is finding the shortest routes to ensure the optimal use of time and resources.

Karim Beguir, co-founder and CEO of UK-based artificial intelligence (AI) startup InstaDeep, told GPU Technology Conference attendees that GPU-powered deep learning and reinforcement learning may have the answer.

The problem with previous efforts is that none of them learn, Beguir noted. “There is no experience which is being gathered from the problems that were solved before,” he said in a blog post. “With the computational investments you’re making to solve this problem, it would be nice if there were some learning.”

Recent work that combines the speed of deep learning neural networks with the deliberate decision making of the Monte Carlo Tree Search technique is leading to breakthroughs that will transform logistics. The AlphaZero program created by DeepMind, which combined neural networks, Monte Carlo Tree Search and the ability to learn from self-play, has created what Beguir called “a new type of AI champion”. AlphaZero is different in that it is not trained on datasets of ideal scenarios - it was given the rules for playing chess and left to deduce how to win at the game, which it did.

“This is happening with zero data,” Beguir pointed out. “The only data is what the system generates by playing against itself.”

Inspired by this advance, InstaDeep is working on injecting Monte Carlo Tree Search into deep learning to apply AlphaZero-like capabilities to solving business problems. Running its models on an NVIDIA DGX-1 AI supercomputer is leading to some powerful developments.

As with AlphaZero, InstaDeep’s AI algorithm does not rely on training with datasets and instead works from scratch, finding progressively better paths between two points. The algorithm also is learning how to pack bins more efficiently, addressing another logistics industry challenge.

Said Beguir: “A few years from now, if your system doesn’t have learnability in it, you’re probably doing something wrong.”

InstaDeep is a member of NVIDIA’s Inception programme, which helps accelerate startups pushing the frontiers of AI and data science.