Reinforcement learning models are trained, using a similar concept by animal researchers to train animals.
Artificial Intelligence technology emulates human behavior. For a very long period, artificial intelligence agents were trained on machine learning models to perform tasks that are usually done by humans. The neural networks of machine learning models are designed and trained in such a format that they perform the tasks without any human intervention or supervision. However, ever since its inception, the researchers and scientists are curious to induce cognitive abilities into artificial intelligence agents.
For a decade, despite the experiments designed to train the artificial neural network by utilizing the human cognitive ability for adopting common sense, the researchers were unable to reach into a reasonable conclusion. The researchers were resorting to behavioral science and neuroscience earlier to induce common sense into the artificial intelligence agents.
But recently, AI researchers from Imperial College London, University of Cambridge and Google Deep mind in a collaborative project were able to induce common sense in the traditional reinforcement learning model using animal cognition. In a paper titled “Artificial Intelligence and the Common Animals”, the researchers cite that it is now possible to apply animal cognition to the reinforcement learning agents in 3D environments to assess the common sense capability of artificial intelligence.
Reinforcement learning involves training the artificial intelligence agents using trial and error method. They perform an array of tasks with repetition and associated rewards. The researchers have taken inspiration from the model which trains animals to perform tasks using rewards, to stimulate common sense into artificial intelligence. Moreover, the challenge pertaining to the existent animal behavior, especially their ability to comprehend the reason to perform a particular task is also understood with the deep reinforcement learning models.
The paper cites that with the advent of reinforcement learning, the assumption about purposeful behavior motivated by the basic needs will be satisfied and the cognitive prowess of an AI system will be evaluated using methods that were designed for animals.
DeepMind has already introduced a combination of deep learning and reinforcement learning. For example, the mid-2010s, reinforcement learning was paired with deep neural networks, which led to the success of Alpha Go, the first program to defeat a top-ranked player at the game of Go.
The paper states that “The Reinforcement learning setting, wherein an agent learns by trial-and-error to maximize its expected reward over time, precludes inactivity and permits any cognitive challenge to be presented by means of a suitably designed environment and reward function.” The reinforcement learning methods are of two types, namely, Model-based and Model-free, which are transition models and maps out the behavior of artificial intelligence agents. In the model-free reinforcement learning method, the agent learns and enacts its policy without reference to a transition model.
Since Deep reinforcement learning model is considered ideal to provoke common sense in the artificial intelligence agents, the training is carried out by using a single task and multiple task settings. In a single task setting, the agent is presented with one task many times, so that its performance can be improved. In multitask, however, the agent learns many tasks together, and the training tasks are often presented concurrently. Moreover, episodes from different tasks are chopped up, interleaved, and stored in a replay buffer, and then presented to the learning component of the agent in random order.
The paper states two criteria’s for training reinforcement learning agents. First, the RL agents should be trained on large suites of varied tasks that involve the opening of container-like objects of various sorts to obtain a reward item. Second, trained agents need to be tested in a transfer setting and the transfer tasks need to be devised that are solvable by the agent.
The researchers state that by using the mentioned approach, the artificial intelligence models will be able to grasp inter-related principles of cognitive behavior and concepts as a systematic whole and that manifests this grasp in a human-level ability to generalize and innovate.