Babak Hodjat

An Adaptive Agent Oriented Software Architecture

Method and agent network architecture for processing a subject message, where each agent has a view of its own domain of responsibility. An initiator agent which receives a user-input request and does not itself have a relevant interpretation policy, queries its downchain agents whether the queried agent considers such message to be in its domain of responsibility. Each queried agent recursively determines whether it has an interpretation policy of its own that applies to the request, and if not, further queries its own further downchain neighboring agents. The further agents eventually respond to such further queries, thereby allowing the first-queried agents to respond to the initiator agent. The recursive invocation of this procedure ultimately determines one or more paths through the network from the initiator agent to one more more leaf agents. The request is then transmitted down the path(s), with each agent along the way taking any local action thereon and passing the message on to the next agent in the path. In the event of a contradiction, the network is often able to resolve many of such contradictions according to predetermined algorithms. If it cannot resolve a contradiction automatically, it learns new interpretation policies necessary to interpret the subject message properly. Such learning preferably includes interaction with the user (but only to the extent necessary), and preferably localizes the learning close to the correct leaf agent in the network.

Introducing an Age-Varying Fitness Estimation Function

We present a method for estimating fitness functions that are computationally expensive for an exact evaluation. The proposed estimation method applies a number of partial evaluations based on incomplete information or uncertainties. We show how this method can yield results that are close to similar methods where fitness is measured over the entire dataset, but at a fraction of the speed or memory usage, and in a parallelizable manner. We describe our experience in applying this method to a real world application in the form of evolving equity trading strategies.

Maintenance of a Long Running Distributed Genetic Programming System for Solving Problems Requiring Big Data

We describe a system, ECStar, that outstrips many scaling aspects of extant genetic programming systems. One instance in the domain of financial strategies has executed for extended durations (months to years) on nodes distributed around the globe. ECStar system instances are almost never stopped and restarted, though they are resource elastic. Instead they are interactively redirected to different parts of the problem space and updated with up-to-date learning. Their non-reproducibility (i.e. single “play of the tape” process) due to their complexity makes them similar to real biological systems. In this contribution we focus upon how ECStar introduces a provocative, important, new paradigm for GP by its sheer size and complexity. ECStar’s scale, volunteer compute nodes and distributed hub-and-spoke design have implications on how a multi-node instance is managed. We describe the set up, deployment, operation and update of an instance of such a large, distributed and long running system. Moreover, we outline how ECStar is designed to allow manual guidance and re-alignment of its evolutionary search trajectory.

Estimating the Advantage of Age-Layering in Evolutionary Algorithms

In an age-layered evolutionary algorithm, candidates are evaluated on a small number of samples first; if they seem promising, they are evaluated with more samples, up to the entire training set. In this manner, weak candidates can be eliminated quickly, and evolution can proceed faster. In this paper, the fitness-level method is used to derive a theoretical upper bound for the runtime of (k+1) age-layered evolutionary strategy, showing a significant potential speedup compared to a non-layered counterpart. The parameters of the upper bound are estimated experimentally in the 11-Multiplexer problem, verifying that the theory can be useful in configuring age layering for maximum advantage. The predictions are validated in a practical implementation of age layering, confirming that 60-fold speedups are possible with this technique.

A Genetic Algorithm to Improve Agent-Oriented Natural Language Interpreters

A genetic algorithm is used to improve the success-rate of an AAOSA-based application. Tests show promising results both in the improvement made in the success-rate of the development and test corpora, and in the nature and number of interpretation rules added to agents.

Tackling the Boolean Multiplexer Function Using a Highly Distributed Genetic Programming System

We demonstrate the effectiveness and power of the distributed GP platform, EC-Star, by comparing the computational power needed for solving an 11-multiplexer function, both on a single machine using a full-fitness evaluation method, as well as using distributed, age-layered, partial-fitness evaluations and a Pitts-style representation. We study the impact of age-layering and show how the system scales with distribution and tends towards smaller solutions. We also consider the effect of pool size and the choice of fitness function on convergence and total computation.

Evolution is the new deep learning

Deep learning (DL) has transformed much of AI, and demonstrated how machine learning can make a difference in the real world. Its core technology is gradient descent, which has been used in neural networks since the 1980s. However, massive expansion of available training data and compute gave it a new instantiation that significantly increased its power.

nPool: Massively Distributed Simultaneous Evolution and Cross-Validation in EC-Star

We introduce a cross-validation algorithm called nPool that can be applied in a distributed fashion. Unlike classic k-fold cross-validation, the data segments are mutually exclusive, and training takes place only on one segment. This system is well suited to run in concert with the EC-Star distributed Evolutionary system, cross-validating solution candidates during a run. The system is tested with different numbers of validation segments using a real-world problem of classifying ICU blood-pressure time series.