| Hong Jin-ho, former professional StarCraft player |
“A variety of factors should be considered in StarCraft ― including resource and unit management or battle tactics ― but the most important thing is how skillfully a player can manage and control hundreds of units at a few major battles,” Korea's renowned former professional StarCraft player Hong Jin-ho told The Korea Times in a telephone interview, Monday.
“Any top-tier professional players cannot control all of them at once, but I think AI programs are capable of doing that, making a huge difference from human beings,” he said.
The remark came two days after Google's AI subsidiary DeepMind announced its partnership with Blizzard's StarCraft II team over opening an open AI research environment by 2017. The open-source application program interface (API) is expected to come as a bigger challenge to an unconfirmed StarCraft player to fight against the next AlphaGo.
“If we (human beings) make enough preparations, however, it may be tough for AI systems to beat us,” he added. “I can say for sure that the human-AI match in StarCraft will also become a bigger challenge to AI systems, compared to the recent historic go battles between AlphaGo and human go champion Lee Se-dol.”
“This is because computers are still more stereotyped in terms of decision-making, even if the recent AI technologies show otherwise,” he said. “Players have to make hundreds of decisions spontaneously in StarCraft, for which computers have yet to overcome humans.”
This past March, the five go matches have made headlines across the world, as it shattered a widespread prediction that the computer-based platform cannot beat humans in such areas where intelligence is highly required. At that time, the world-class go player acknowledged his total defeat against the AI platform.
Following the matches, the DeepMind team announced its next challenging partner ― StarCraft ― which has left a big footprint in real-time strategy game history. DeepMind welcomed its partnership with Blizzard, as the StarCraft series has been widely recognized as the “pinnacle of 1 versus 1 competitive games.”
“StarCraft's high-dimensional action space is quite different from those previously investigated in reinforcement learning research; to execute something as simple as ‘expand your base to some location,' one must coordinate mouse clicks, camera and available resources,” DeepMind said in a statement. “This makes actions and planning hierarchical, which is a challenging aspect of reinforcement learning.”