Essentially, the aim of artificial intelligence (AI) is to achieve a state where a computer can learn, react & differentiate on an equal scale when compared with a human being. Often computers fare better than humans when all known parameters in a given environment are programmed in, the computer can generate every option at lighting speed, & come up with the best answer. However, when it’s impossible to input every possible scenario, computers fare less well.
Take, for example, a game of chess. All the best moves possible in a given situation are programmed into the machine, which makes it an almost unbeatable opponent, since it can assimilate every known option at anytime. Humans, by contrast, may forget certain winning moves, that’s part of what makes us human – it, also, gives the machine its advantage in these situations.
Games have long been a method for testing & developing AI, but only recently has the focus been on games which require more than a predefined set of outcomes to achieve a result. Facebook’s AI research team, the Lorraine Research Computer Science Laboratory & the University College London, have embarked on a study to test AI reactions in an old fashioned text-based game with hundreds of thousands of unknown parameters, the aim of which is to get a computer to learn from experience, rather than to simply select options from a pre-taught result set.
To this end the team has created an AI learning environment in the form of a massive text-based adventure RPG(Role Playing Game) called, ‘Light’. The game has been crowdsourced, & gives the scientists an opportunity to examine the effect of ‘grounding dialogue’ its AI participants are capable of. Grounding dialogue, basically means a specific context based exchange between characters in a particular or unprepared situation. This is the ultimate frontier for AI, & the hardest boundary to overcome.
The game has crowdsourced back stories from hundreds of people with the aim of creating scenes which require natural language & natural (human) intelligence to navigate. Currently, humans have contributed a total of 663 locations, 3,462 objects & 1,755 characters to Light. Furthermore, humans have structured 10777 situationally based actions, dialogues & object based reactions that researchers can use to train AI players.
The game is presented in such a way that both human & AI participants can play, interact & learn. Technically, the system uses Facebook’s PyTorch machine learning framework to gauge context based reactions & score AI participants. Google’s Bidirectional Encoder Representations from Transformers (BERT) has, also, been engaged to build a comparison between context & response.
Predictably, the more information the machine is given about the environment, possible actions, dialogue & acceptable responses the better it fairs, but the entire concept & development promises to demonstrate the degree of learning AI is capable of at its present state of development. An experiment the world is very keen to follow.
