There have been several advancements in the AI field, particularly with the use of machine learning and neural networks. Some of these approaches are categorised under supervised or unsupervised techniques. An alternative method is reinforcement learning, an area of AI research where intelligence is built on rewarding actions, akin to how children learn the difference between good and bad behaviour.

Reinforcement learning uses neural networks and deep learning techniques to maximise the the long term value of a desired goal that is reached over time. In order to do so, rewards are used to excite the reinforced agent in a scenario to perform better during its mission to achieve the goal. This is done by tracking the agent’s state in the environment and pairing it with rewards. The rewards can be either negative or positive. A negative reward will ensure that over time the agent will seek to perform the undesired action less often or not at all, while a positive reward will enforce the agent to repeat this behaviour for a desired outcome. The policy is the strategy used by the agent to get the most rewards. It enables reinforcements learning to pair actions with highest reward outcomes.

A recent release of MLAgents, a toolkit developed by Unity Technologies for reinforcement learning experiments in games, has allowed faster deep learning simulations on a single machine while acting as an enabler for both playable and non playable character behavioural modelling. Snoopy Pop, a bubble shooter game developed by JamCity was used as a platform for testing this new update. The developers at JamCity managed to create an agent that is able to play many level of the game with increased difficulty. This was possible to environment parallelism. In total, sixteen environments were used to speed up the training process for the agent. The trained brain receives input from each instanced simulation to gather more information from the environment at the same time.

For more information on this research, check out

Also visit for reinforcement learning definitions and further reading.