Spatial Representation in Biological and Artificial Autonomous Agents
Rats can orient themselves in an environment; they easily learn to navigate in an environment in order to approach a goal (food or escape platform). How is this possible? In the hippocampal formation of the rat, cells have been found that code for the current position of the rat (place cells) and for its orientation (head direction cells). These cells are thought to be a substrate for the representation of the environment and serve as a basis for spatial learning and navigation.
Autonomous agents must perform similar non-trivial navigation tasks in unpredictable environments, such as obstacle avoidance, homing, search for goals, in ways that remind the observer of animal behaviour. There are two opposite views of how such navigation tasks can be solved computationally. One approach splits the navigation task in two subtasks, representation and action. According to this approach, taking actions will be rather simple if a reliable model of the environment is available to the agent. The main problem consists in designing such an internal representation and making available to the agent in order to achieve a predesigned goal. The other approach makes no use of an explicit representation of the environment. Instead, a combination of simple rules and reflexes provides the agent with a set of intelligent behaviours. Here, the major challenge is that of devising an efficient methodology for automatic development of appropriate connections and priorities among several simple low-level processes.
Our approach is inspired by the hippocampus of the rat and can be situated somewhere in between these two extremes. The rat hippocampus is, to a certain degree a spatial representation. In contrast to representations in robotics, it is highly redundant, it has broad overlapping place fields, and and it has no topological organization.
Starting from a model of the rat hippocampus we ask the following questions:
To answer these and other questions we use neural modeling together with implementations on the Khepera Robot. Neural Network approches are applied in order to generate learning and adaptive behavior. Specifically, during the initial exlporation of the environment, the robot develops 'place fields', viz., units which are activated only if the robot is at a specific location in the environment. Given the place fields, simple learning rules like asymmetric Hebbian learning or Reinforcement learning can be applied to solve, e.g., homing tasks.
This project finished in 2008:
Sheynikhovich, Denis ; Chavarriaga, Ricardo ; Strösslin, Thomas ; Arleo, Angelo ; Gerstner, Wulfram (2005)
T. Strösslin, D. Sheynikhovich, R. Chavarriaga, and W. Gerstner (2005)
T. Strösslin, R. Chavarriaga, D. Sheynikhovich, and W. Gerstner (2005)
Ricardo Chavarriaga, Thomas Strösslin, Denis Sheynikhovich and Wulfram Gerstner (2005)
Ricardo Chavarriaga, Thomas Strösslin, Denis
Sheynikhovich and Wulfram Gerstner (2005)
A. Arleo and W. Gerstner (2000)
For additional references, consult the list of Publications of the lab
Please send comments on this page to: email@example.com