Science Journey - Smarter AI for a More Sustainable World: Building an AI Model You Can Trust

Chris Yeh
Computing and mathematical sciences graduate student

Download a printable packet about Chris's presentation, including vocabulary and classroom activity.

About the Presentation

Artificial intelligence • Computer science • Sustainability

Sustainability is the challenge of meeting the needs of the present without compromising the ability of future generations to meet their own needs. Solving the world's biggest sustainability challenges—like eliminating extreme poverty, protecting wildlife, and making clean energy more reliable—often means making tough decisions with limited information. For example, where should food and medical aid go to help the most people? How many endangered animals are really left in the wild? How much electricity will people need next week, or next year?

Scientists like Caltech graduate student Chris Yeh are using artificial intelligence (AI) to answer many of these questions and make predictions about what will happen to Earth's resources in the future, so that everyone, from electric companies to farmers to lawmakers, can make smarter sustainability decisions. But there's a catch: Today's AI systems generally cannot tell us how confident they are in their predictions. That becomes a problem when high-stakes decisions, like where to send aid or when to turn on or off power plants, depend on getting it right.

Chris works on building AI systems that not only make predictions but also say how certain or uncertain those predictions are. Quantifying uncertainty helps scientists, conservationists, and engineers understand when they can trust AI and when they should be more cautious about using the information AI provides. By designing AI to work better with uncertainty, we can help make smarter, safer choices for a more sustainable future.

About the Speaker

Credit: Courtesy of Chris Yeh

Chris Yeh is a PhD student in computing and mathematical sciences at Caltech, co-advised by Professors Yisong Yue and Adam Wierman. His research focus is on artificial intelligence (AI) algorithms that combine uncertainty quantification and decision-making, particularly with applications in sustainability and energy systems. Growing up in Southern California, he first learned to program computers by designing websites, and he explored his love for the outdoors through camping and backpacking trips throughout California. Chris received bachelor's and master's degrees in computer science from Stanford University in 2019. He found his passion for using AI to solve sustainability challenges as part of Stanford's Sustainability and AI Lab. Outside of his research, Chris plays ultimate frisbee on the Caltech Aftermath club team and is a cellist in the Caltech Orchestra.

Enrich your knowledge around the lecture topic by reviewing relevant terms, provided by Chris.

Adapted from "Machine learning with matchboxes," by Pavel Anni and from Wisconsin Materials Research Science and Engineering Center Education and Outreach Group (University of Wisconsin–Madison).

This activity teaches the basics of machine learning using craft supplies—no computer needed! It is based on a game called Hexapawn, played on a 3×3 board with three white chess pawns (or other game pieces) and three black pawns (or other game pieces).

GAME RULES

• Pawns move forward one space.
• Pawns capture opponents by moving diagonally forward.
• You win if:
  • One of your pawns reaches the opposite side of the board,
  • You capture all opponent pawns, or
  • Your opponent has no moves left.

WHAT YOU NEED

• 24 paper lunch bags
• 6 game pieces (3 of one color, 3 of another)
• About 50 beads or buttons in 4 different colors
• Paper and markers that match the colors of the beads
• Scissors, glue
• A small 3x3 square board and game diagrams; draw your own or download a printable version from the University of Wisconsin-Madison

HOW TO BUILD THE "LEARNING MACHINE"

For the purposes of these instructions, three game pieces are black, three are white. You will teach a simple robot to play as black.

1. Create paper bags for robot decision making: Your 24 bags each represent a board layout (i.e., where each black and white piece is on the board) in which it is black's turn to choose the next move (the second, fourth, and sixth turn of the game).
2. On the front of each bag, draw or cut and paste a game diagram. Use colored markers to draw arrows indicating each move available to the black pieces in that scenario. You will have two bags for the robot's second turn, 11 bags for the fourth turn, and 11 bags for the sixth turn.
3. Inside each bag, place colored beads that correspond with the colored arrows on the front of the bag. Each bead represents a "choice."

HOW TO PLAY

1. You play white. Move your pawn.
2. Identify the paper bag scenario that matches the way the pieces are now laid out on the board (or the mirror image of the layout).
3. Shake the bag and pick one bead at random.
4. Make black's move based on the bead color, then put the bead back in the bag.
5. Continue up to six moves until someone wins.

The robot "learns" by playing multiple games.

• If the robot (black) loses, remove the bead that determined the last move.
• If the robot's last bag is left with only one bead, and the robot still loses, remove both the bead from the final bag and the bead that determined the move prior to that one.
• Over many (30-plus) games, the robot keeps only "good" moves and eventually becomes very strong—even unbeatable.