A genetically evolved neural network for an action selection mechanism in behaviour-based systems

The world of robotics has grown so much that it has reached a state where it can be trusted with many real-world applications, especially those that involve a high safety risk for human effort. From its ‘humble’ beginnings operating in astatic and controlled environment, robot platforms are now required to operate in dynamic and unknown environments, which traditionally require human intelligence for real-time decision making. Development of a control system for a robot platform to handle such scenario can be very demanding. The system requires complex decision making capabilities in order to be sufficiently robust and responsive to the dynamics of its environment.

One possible approach is to implement a behavior-based system, which ‘reacts’ to its environment rather than using preprogrammed rules of engagement. However, other than the accuracy of its sensors, the success of a behavior-based system relies largely on its Action Selection Mechanism (ASM) module, which is basically a behavior coordination method. Common implementations of behaviour coordination method can be categorised into two: arbitration and command fusion. Consequently, deciding on a suitable coordination method for a particular task in an unknown environment presents a similar complex issue. To handle this, the more popular approach is to use Artificial Intelligence (AI) in the development of ASM modules.

In this thesis, a Genetic Algorithm (GA) has been used to evolve a neural network engine that is used as an ASM module for a behavior-based system. The proposed control architecture implements a basic GA to train the synaptic weights of a simple Multi-Layered Perceptron (MLP) feed-forward Artificial Neural Network(ANN) in identifying a suitable formulation of ASM. A simple, yet found to be sufficiently adequate, fitness function has been formulated in order to ensure the effectiveness of a GA in evolving the system. The proposed fitness function is defined as such that it can be generalised and applied to any robot control tasks. The proposed system has been tested using simulation software in two common robot mission scenarios involving unknown environments: search and exploration, and target tracking.

Simulation results show that the proposed Genetically Evolved ASM(GEASM) can dynamically manage the behavior coordination method that enables the system to achieve mission objectives in both test scenarios. For the search and exploration mission, the GEASM managed to achieve a 93% success rate compared to other architectures, with the nearest competitor at 67%. As for the target tracking mission, the GEASM achieved a stunning 100% success rate, compared to the next best at 75%. Since the test environment is actually different from the one used in training the proposed system, it can be projected that the GEASM can actually enable a system to perform in an unknown environment with a significantly high probability of success.