Such traps are represented as cul-de-sacs that are close to the final destination, but that can not be escaped without a step back, away from the ultimate goal at least temporary. In such environments, objective-based solvers basically can not find optimal solution or any solution at all, because they do not have the necessary internal machinery for committing the leap-of-faith and move backward from the target in order to eventually find a way out.
Google will beat Apple at its own game with superior AI.
The objective-based solvers depends on the best attempts of their designers to assess the operating environment and develop a better way to achieve the ultimate goal. But, as it happens in the real world, preliminary assumptions often can not account for all the traps on the way to the goal because of the extreme complexity of the environment settings. And even in simple artificial environments, such as maze navigation, it often happens that objective-based solvers can not find the optimal solution within the adequate execution time and computational resources allocation.
But for successful autonomous execution in real world environments, it is critically important to create Intelligent Agents capable of quickly finding a solution with minimal computational load. In the experiment we studied how Novelty Search NS method of fitness function optimization performs compared to traditional objective-based ones for unsupervised training of Artificial Intelligent Agents to do spatial navigation in complex maze environment.
The main idea behind NS optimization is to rather look for novel outcomes in the search space than the distance to the final objective: the maze exit. The Novelty Search assigns higher fitness values to the Intelligent Agent capable to find the most novel solution among all previous tries. Despite its ignorance to the final objective the NS happens to be extremely effective optimization method capable of breeding AAIA, which crack deceptive real-world tasks even in the realms where traditional objective-based methods have failed completely. The main assumption about what makes this possible, is that in order to reach final goal, AAIA must find several intermediate goals stepping stones which in most cases do not resemble the ultimate objective.
Sometimes Intelligent Agent must step back to avoid deceptive traps. By doing this it will see a decrease in value of objective-based fitness function for a moment but will get a better outcomes in the future.
This is one of the fundamental properties of the real-world environment that the exact route to the final objective in most cases can not be predicted in advance, and all intermediate stepping stones should be found by taking the path. In the experiment we combined NS with NeuroEvolution of Augmenting Topologies algorithm which efficiently evolve ANNs through complexification by augmenting its topologies. Autonomous Artificial Intelligent Agent, designed to solve the maze, has ten input sensors that allow collecting information about the environment and two output effectors controlling its movements through the maze see Figure 1.
The final objective of the agent is to go through the maze and find a way out. The input sensors are: six range finders that indicate the distance to the nearest obstacle blue arrows and four pie-slice radar sensors slices of red circle that act as a compass towards the goal maze exit , activating when a line from the goal to the center of the agent falls within the pie-slice. The green-yellow arrow in the center points to the movement direction of the agent. The agent is also equipped with two effectors producing a forces that respectively turn and propel the robot, i.
The configuration of the seed genome of the solver agent can be summarized as follows see Figure 2 :. The input neurons has following numbers on the diagram above:. The Novelty Search optimization method is based on novelty metric calculation for each solver agent after performing a certain number of time steps in simulation of maze navigation for that agent.
The novelty metric biases the search in a fundamentally different way than the objective-based fitness function which depends only on the distance from the agent to the exit and determines the behavior-space through which the search will be performed. Therefore, since what is important in the maze, this is where the solving agent ends navigation, then for the maze domain, the behavior of a navigator is defined as its final position.
The novelty metric then maximizes the N-nearest neighbor distance between the final positions of all known solving agents, i. The effect of this novelty metric is to reward the solver agent for ending in a place where none have ended before and the method of traversal is ignored. This measure reflects that what is important, is to reach a certain location i.
Thus, although the novelty metric has no knowledge of the ultimate goal, a solution that reaches the goal can appear novel.
Search Algorithms in Artificial Intelligence - By
In addition, the comparison between fitness-based and novelty-based search is fair because both scores are calculated only based on the distance of the final position of the agent from other points. As for deceptive environments we choose two types of maze environments with different complexity as was recommended in [ this research ]: medium and hard maze. The maze configurations was designed in such a way as to create many cul-de- sacs with strong local optima, deceiving the objective-based optimization methods.
The Medium Complexity Maze Environment.
The first experiment to establish baseline performance metric was performed using maze configuration of medium complexity. The Novelty Search optimization was combined with Neuro-Evolution of Augmented Topologies algorithm which use genetic neuro-evolution process to create a population of organisms capable of solving a maze.
We also compared its performance with the objective-based optimization method for the NEAT algorithm, where fitness function optimization was dependent on how close final destination of produced organism is from the exit of the maze. The final performance metric of each Autonomous Agent created for both optimization methods depends only on how close to the maze exit is the final destination of the solver after stimulation steps.
Problem Solving and Search in Artificial Intelligence
Thus, despite the various methods of fitness function optimization, the final results can be compared for both methods. Each experimental trial was performed with 2 epochs of evolution or until a winner is found. Applying Novelty Search based optimization it was possible to get the winner in 10 form 10 trials with optimal genome found approximately within 50 generations. An Artificial Neural Network produced by an organism with a near optimal genome has 16 neurons with only three hidden units, i.
And it is able to control maze solver agent with a spatial error of about 1. Among with the two additional hidden units neurons , the recurrent link was developed at the output neuron 13 angular velocity effector — see Figure 3.
It is also interesting to consider the hidden neuron 91, which seemingly have learned the complex behavior of the steering to the exit of the maze when it is discovered to the right or behind the agent. The hidden neuron connected with input sensor 11 radar sensor: RIGHT , has learned to influence the steering of the agent in the direction of the exit of the maze, since most of the time the exit is on the right bottom relative to the agent.
The hidden neuron 12 which is introduced in seed genome operates as main control-and-relay switch relaying signals from sensors and other hidden neurons to the effectors neurons 13, On the Figure 4 presented a diagram of the maze solving simulation by solver agents controlled by ANNs, derived from the genomes of all organisms introduced into the population until a winner is found.
Agents are coded by color depending on which species the source organism belongs to. The fitness of agent is measured as the relative distance between its final destination and maze exit after running simulation for certain number of time steps in our setup. The initial agent position is at the top-left corner marked with green circle and maze exit at the bottom-right marked with red circle.
The results are presented for an experimental trial producing the configuration of the winner genome depicted at Figure 3. Applying objective-based optimization , it was possible to create the winners capable of solving medium maze configuration in 9 from 10 trials. Problem solving and search is a central topic in Artificial Intelligence. This course presents several techniques to solve in general difficult problems. The goals can be acquierd by studying the lecture material, solving the exercises of the tutorials and developing an implementation for a practical problem.
Subscribe to events of this course icalendar. TU Dresden. Account Log in. Edit tools Discussion View source History. Suche Search. Jump to: navigation , search. Prerequisites Basic knowledge of theoretical computer science and Logic. Organisation The goals can be acquierd by studying the lecture material, solving the exercises of the tutorials and developing an implementation for a practical problem.
The practical work should be performed in groups of two students throughout the semester with regular updates on the progress. Stuart J. Russell and Peter Norvig.
Related Search in Artificial Intelligence
Copyright 2019 - All Right Reserved