DeepMind and the Development of Artificial General Intelligence
First Chess and now Go. What’s the next to go?
DeepMind is a leader in artificial intelligence research most famous for its AlphaGo program, an AI program that defeated the world鈥檚 best player at Go[1]. Its sole mission is to 鈥減ush the boundaries of AI, developing programs that can learn to solve any complex problem without needing to be taught how鈥�[2]. As a result, it aims to be at forefront of AI research, so that it can develop and apply artificial intelligence to a wide range of disciplines. It tackles real-world problems by collaborating with experts on long-term research projects centered on the application of AI to different fields.
Despite recent advancements in AI, researchers at DeepMind recognize that there are constraints and areas of artificial intelligence that have yet to be explored in detail. For example, many of the problems tackled by AI have had limited action states, perfect information between participants, or have had one agent acting at a time. In the real world, most of these constraints don鈥檛 exist and actions are more dynamic and varied. As a result, DeepMind has teamed up with Blizzard to use Starcraft II (a real-time strategy game) as a research learning environment, because gameplay in Starcraft II represents a 鈥渕ore difficult class of problems than considered in most prior [AI] work鈥�[3] because it removes many of the constraints mentioned above. 聽In addition, Starcraft II brings to the forefront an issue that AI, and humans, have difficulty dealing with: uncertainty and lying. A key component of Starcraft gameplay is the idea of a 鈥渇og of war鈥�. Each player is only able to see parts of the map where they currently have a unit[4]. This lack of visibility increases the amount of uncertainty in a game and creates the opportunity for players to lie and provide false information[5]. As a result, machines cannot simply 鈥渋terate鈥� and make the best move but must make a judgment based on the incomplete set of information that it has. This, in comparison to having perfect information, is much closer to reality for many of the problems that AI is trying to tackle.
DeepMind鈥檚 goal with Starcraft II is to build 鈥渁rtificial general intelligence鈥� by better understanding what the 鈥渓earning paradigm鈥� is, so they can build an agent that can learn to 鈥減lay any game without much prior knowledge鈥�[6]. Unlike previous AIs like Deep Blue (chess) or IBM鈥檚 Watson, newest AI by DeepMind is meant to 鈥渞emember, to strategize, and to learn鈥�[7]. 聽Since DeepMind understands that their overall goal is lofty, they have broken the problem down into 鈥渟maller, more manageable chunks鈥�[8], something that makes games a particularly attractive research environment. They鈥檝e partnered with Blizzard to gain access to a database of replays that the AI can learn from and have designed mini-games to break the overall gameplay into smaller problems[9]. For the foreseeable future, DeepMind is trying to tackle the big picture problem by looking at the little pieces with the promise of general application in the future. They have opened the resources to the public to generate interest and gain insight from third parties. I believe DeepMind should also apply their AI in more practical ways in parallel with their work with Starcraft, so they can identify some transitional difficulties early on. They don鈥檛 need to address any of the issues before the AI is more developed, but it could be helpful to see how the AI performs outside of the controlled environment if it is meant to have general applications.
While I understand the pursuit of the overall big picture for building artificial general intelligence, does this grand quest make sense in a world with constrained resources? Will we, in the foreseeable future, use AI in a way where constraints will be minimal, if not nonexistent? Does it make more sense to tackle a problem with more constraints if that鈥檚 the intermediate step anyway?
Word Count: 800
Photo Credit: DeepMind
[1] Elizabeth Gibney, 鈥淲hat Google鈥檚 winning Go algorithm will do next鈥�, Nature (Mar 15, 2016).
[2] DeepMind website 鈥淎bout Us鈥�, https://deepmind.com/about/
[3] Oriol Vinyals et al, 鈥淪tarcraft II: A New Challenge for Reinforcement Learning鈥� p. 1, Cornell University (Aug 2017).
[4] Hochul Cho, 鈥淚nvestigation of the Effect of 鈥楩og of War鈥� in the Prediction of StarCraft Strategy Using Machine Learning鈥� p. 2, Computers in Entertainment (Jan 2017).
[5] Jonathan Cheng, 鈥淗umans Still Rule in This Game 鈥� To win 鈥楽tarCraft,鈥� machines need to learn to lie鈥� p. 2, Wall Street Journal (Apr 23, 2016).
[6] Justin Groot, 鈥淐hecking in with the DeepMind Starcraft II Team: Interview with Oriol Vinyals鈥�, Blizzard News (Mar 4, 2018).
[7] Christina Beck, 鈥淣ext AI Challenge: Computers take on StarCraft鈥�, The Christian Science Monitor (Nov 05, 2016).
[8] 鈥淪hall we play a game?; Artificial intelligence鈥�, The Economist (May 13, 2017).
[9] Oriol Vinyals et al, 鈥淪tarcraft II: A New Challenge for Reinforcement Learning鈥� p. 1, Cornell University (Aug 2017).
It is definitely interesting to see the parallels you drew between DeepMind’s work with Starcraft II and projects like Deep Blue and IBM’s Watson, as although Starcraft II is still just a game, it seems to represent a natural progression and the next evolution of AI. With constrained resources, you certainly have a point that maybe this technology could be applied somewhere where social impact could be better felt, but I think that as each iteration of AI progresses from one challenge to the next, we get closer to principles and capabilities that are more generalizable for real world applications. As an added bonus, Starcraft II may be able to draw more data scientists who are attracted to this project and its unique and fun learning environment. Great article!