Learning to Search from Demonstration Sequences | Dixant Mittal

💬 Discord: https://discord.gg/h8NVzwnysW 💻 GitHub: https://github.com/centaurinstitute 🤝 LinkedIn: https://www.linkedin.com/company/centaur-ai-institute/ 📢 New Initiative: Neuro-Symbolic Agentic Protocol ⭐ Star us on GitHub https://github.com/centaurinstitute/neuro-symbolic-agentic-protocol — ​​​📝 Abstract: Search and planning are essential for solving many real-world problems. However, in numerous learning scenarios, only action-observation sequences, such as demonstrations or instruction sequences, are available for learning. Relying solely on supervised learning with these sequences can lead to sub-optimal performance due to the vast, unseen search space encountered during training. In this paper, we introduce Differentiable Tree Search Network (D-TSN), a novel neural network architecture that learns to construct search trees from just sequences of demonstrations by performing gradient descent on a best-first search tree construction algorithm. D-TSN enables the joint learning of submodules, including an encoder, value function, and world model, which are essential for planning. To construct the search tree, we employ a stochastic tree expansion policy and formulate it as another decision-making task. Then, we optimize the tree expansion policy via REINFORCE with an effective variance reduction technique for the gradient computation. D-TSN can be applied to problems with a known world model or to scenarios where it needs to jointly learn a world model with a latent state space. We study problems from these two scenarios, including Game of 24, 2D grid navigation, and Procgen games, to understand when D-TSN is more helpful. Through our experiments, we show that D-TSN is effective, especially when the world model with a latent state space is jointly learned. ​🔗 Link to paper: https://iclr.cc/virtual/2025/poster/27934 ​🔬 Bio: Dixant Mittal is a Lead Data Scientist at Paytm, where he is actively involved in the development of small language models and generative AI products at scale. He holds a Ph.D. in Computer Science from the National University of Singapore, where he was advised by Professor Wee Sun Lee as a member of the Learning, Inference and Planning Group. His doctoral research focused on combining learning with planning algorithms. ​Prior to Paytm, he worked as a Research Scientist at Moovita, developing foundation models and decision-making systems for autonomous vehicles. He also completed his master's degree at NUS under the supervision of Professor David Hsu, with a thesis on active information gathering for visual grounding. ​His research interests span machine learning, reinforcement learning, planning and search, graph learning, and approximate inference. — 🗨️ ​Join us for an interactive session exploring Neuro-Symbolic AI, the emerging paradigm that blends the strengths of neural networks with symbolic reasoning. We will discuss how hybrid approaches can enhance generalization, interpretability, and reasoning, and how these methods are shaping the future of intelligent systems. Whether you’re a researcher, engineer, or simply curious about the cutting edge of AI, you’ll find an engaging space to learn, connect, and exchange ideas. Event Format: 🔍 Reading Group 🎙️ Panel Discussion 🧩 Tutorials/Workshop Series 💬 Research Roundtable 🛠️ Open-Source Project Review 📄 Paper Pitch 🤝 Collaboration Hour 🍵 Networking Mixer #AI #NeuroSymbolic #FutureOfAI