Behavioural cloning will push the limits of reinforcement learning – The Times of India Blog

Behavioural cloning (Imitation learning) has been there in human race since ages. For learning various skills, we try to imitate the actions of the experts in that field. To become better sportsperson in a particular game, new entrants watch the videos of Olympic winners and try to imitate their actions. Similarly, children follow the actions of their parents while watching them throughout the early stages of their life.

Reinforcement learning is based on the philosophy of evaluating all the state space possibilities in context of agent and reward paradigm. While it has shown tremendous potential in solving many problems, but it has its own limitations in complex environments. Many times, it is not possible to define the state and reward due to the fluidity of the situation and you are not able to divide the activity into discrete steps or states.

Recent experiments by researchers have shown that if we can combine behavioural cloning with reinforcement learning mechanism, then it gives more promising outcomes. Behavioural cloning is kind of a supervised learning technique where you are pre-labelling some of the action sequence as best practices. In this the system will have a faster learning curve.

There are tremendous applications which will be suitable for imitation learning algorithms.

Improving our work force: We can train our AI system on the videos of best employees in your work force and can then benchmark others based on that. For example, in armed forces, if soldiers are supposed to respond in a certain way in specific situations, then the system will be trained on the best soldiers responding to those situations. These videos frames for best soldiers actions can be labelled for different activities. Rewards and penalties can be described based on these actions. Now the videos of any other soldier can be taken responding to those situations and AI can be applied to determine the final performance score. This kind of evaluation is applicable in any domain where some physical activity/skill is involved in completing the task. It will help you to improve your work force and eventually increase your productivity. Even if you want to automate your workplace then Robots can be trained using imitation learning. Instead of making the Robots to learn from zero knowledge base, they can learn from the best practitioners of the field using behavioural cloning.

Imitation Learning: As the Robots will become the norm in day-to-day life, it will be essential to localize them as per the cultural, ethical, and social requirements of the community. In such cases, imitation learning will be very useful. These machines can be augmented using the videos of local representatives on various customs, salutations, routine conversations, gestures etc and the robots will learn to respond accordingly in addition to their core task for which they have been built.

Physically Challenged People can operate devices: Equipment like TV, Washing Machine, Kitchen Equipment, Gardening equipment and others can be easily trained to be used by physically challenged people based on specific actions through voice, hand gestures, eye gestures etc depending upon the abilities of that person.

Autonomous cars: Similarly Autonomous cars can be trained using behavioural cloning of expert drivers. Many companies are already at advance stages of research on this to be used for their vehicles in specific terrains.

Eventually, the punchline is that if we have an example of some best action practices being done by animals, birds, humans, machines or other mechanisms, these actions can be recorded and labelled.

Then the AI models can be trained using the combination of Imitation learning and Reinforcement learning for the AI systems to improve their performance/accuracy.

Views expressed above are the author's own.

END OF ARTICLE

More:

Behavioural cloning will push the limits of reinforcement learning - The Times of India Blog