The Evolution of AI Game Players: From Chess to Go to the Future
Table of Contents
The famous 1997 chess match between IBM's Deep Blue and world champion Garry Kasparov represented a historic achievement for artificial intelligence. For the first time, a computer system was capable of defeating the greatest human chess player under normal tournament rules. But Deep Blue's brute-force approach relied heavily on processing power rather than intelligence. Two decades later, Google DeepMind's AlphaGo system took a completely different approach to conquering the ancient Chinese game of Go. By combining neural network learning and tree search techniques, AlphaGo displayed true intuition and strategic mastery. Its stunning 4-1 victory against legendary Go master Lee Sedol in 2016 was an eye-opening demonstration of how far AI capabilities had advanced.
So how did we get to this point? Let's look back at some of the key milestones that have led to the current state of the art in game playing AI.
 |
| The Evolution of AI Game Players: From Chess to Go to the Future |
Origin Stories: Early Days of AI Game Play
- Back in the 1950s, rudimentary chess programs were developed that could think a few moves ahead by following basic rules of the game. They calculated the best next move within a very limited scope.
- By the 1960s, AI researchers were making some strides in game AI with more advanced programs. Arthur Samuel created a checkers playing program that actually improved through a primitive form of machine learning. It could teach itself by playing against modified versions of itself.
- In the 1970s and 80s, neural networks and reinforcement learning techniques were applied to games like backgammon. Programs like TD-Gammon honed their skills through repeated self-play. The more they played, the better they got.
- Hardware constraints of the time meant these early game AIs were still fairly basic compared to what we have today. But important foundational work was being done applying new AI methods to game environments.
Stepping Up to Grandmaster Level
IBM's Deep Thought system was the first to defeat world-class chess players in tournament play in the late eighties. Building on these capabilities, Deep Blue could consider up to 200 million positions per second by leveraging parallel processing power. This computational muscle enabled it to look many moves ahead and beat Garry Kasparov, considered the greatest chess talent ever. However, some argued that Deep Blue succeeded mainly through brute computing force rather than higher-level understanding of the game. Its strategy was still dictated primarily by human chess knowledge programmed into the system.
 |
| The Evolution of AI Game Players: From Chess to Go to the Future |
Defeating Humans at Their Own Game
The next generation of game AIs aimed to demonstrate deeper comprehension of these games through more human-like intuition and learning capabilities. IBM Watson's triumph on Jeopardy! in 2011 involved complex natural language processing to parse tricky wordplay-laden clues. Google DeepMind's AlphaGo used neural networks to first learn the game by analyzing millions of humans Goes moves. It then honed its skills further by playing against more advanced versions of itself through reinforcement learning. Unlike chess, Go has an enormous possibility space and deep strategic subtleties. AlphaGo's ability to teach itself to mastery purely from gameplay experience, developing unconventional moves that shocked expert Go players, proved that game AI had entered a new paradigm.
Ongoing Challenges in Game AI
- Real-time strategy games like StarCraft remain challenging for AI due to hidden information and huge decision complexity.
- Multiplayer collaborative games require skills like communication, teamwork and coordination that AI still struggles with.
- Areas like abstract reasoning, spatial cognition and handling uncertainty come naturally to humans but are difficult for current AI.
- Transferring game proficiency to real-world applications remains limited, though games provide useful steppingstones to address simplified versions of real-world problems.
 |
| The Evolution of AI Game Players: From Chess to Go to the Future |
The Future of AI Game Play
As algorithms and computing power continue advancing, AIs are expected to surpass humans in additional game genres, including multiplayer competitions. Reinforcement learning and self-play should enable AIs to become competitive without needing extensive human gameplay data. Cloud-based game AI services could allow developers to easily incorporate intelligent agents into new games. And personalized NPCs with adaptable difficulty could provide customized challenge levels and gameplay styles for each human player. While games provide useful milestones for AI capabilities, the ultimate goals are generalized intelligence and transferring these advances to benefit other realms like science, education, business, and more to improve human lives.