Method and device for the control of an autonomously exploring robot

• US 5,774,632 A
• Filed: 11/13/1995
• Issued: 06/30/1998
• Est. Priority Date: 11/15/1994
• Status: Expired due to Fees

• Current Assignee: Alexander Kaske
• Original Assignee: Alexander Kaske
• Inventors: Kaske, Alexander
• Primary Examiner(s): Downs, Robert W.
• Assistant Examiner(s): SHAH, SANJIV
• Application Number:
  US08/554,948
• Time in Patent Office: 960 Days
• Field of Search: 395/22, 395/23, 395/24, 395/10, 395/81, 395/94, 395/88, 701/23, 364/167.01, 382/103, 318/601
• US Class Current: 706/25
• CPC Class Codes: B25J 9/161 Hardware, e.g. neural networks, fuzzy logic, interfaces, processor
• CPC Class Codes: B25J 9/1694 characterised by use of sensors other than normal servo-feedback from position, speed or acceleration sensors, perception control, multi-sensor controlled systems, sensor fusion
• CPC Class Codes: G05B 2219/33012 Kohonen network, single layer with neurodes, associated with codebook vector
• CPC Class Codes: G05B 2219/33022 One network for learned signal values, one network for unknown signal values

Abstract

In a control method and a control device for an autonomously exploring robot, sensory values induce, in a strategy generating network, control sets according to inner connectivity which is acquired in a training phase. The control sets control acting units which act on the environment. Control sets and sensory values are given as input to a network system which distinguishes between already trained pairs of values and unknown pairs of values.

Claims

1. A method for controlling an autonomously exploring robot comprising the steps of:

capturing values, using a sensory device, which are fed as input values into an artificial neuronal strategy generating network,

generating control values with the strategy generating network, according to input and trained synapse strengths, which control acting units of the robot,

feeding the values captured by the sensory device and the control values generated by the strategy generating network into an artificial neuronal network system including a plurality of further networks,

storing said input values of the sensory device and corresponding control values, which constitute pairs of values, into a short term memory,

distinguishing between pairs of values learned during a training phase and unknown pairs of values with a dynamic analyzer on a criterion of dynamic stability of the network system,

extending an initial repertoire of said strategy generating network and the network system by assimilation of selected values captured by the sensory device and the control values generated with the strategy generating network in a training process when the dynamic analyzer identifies a pair of values as known and belonging to a successful string of values,

selecting unknown pairs of stored values which precede the known pair of values,

training the network system until adaptation of synapses strengths of the network system results in a stable output pattern,

feeding the strategy generating network with the stable output pattern, and the selected unknown pairs of values, and

training the strategy generating network, until all selected pairs of stored values are trained, thereby extending the initial repertoire of the strategy generating network and network system.

2. A method according to claim 1, and further comprising the steps of:

utilizing a recursive learning scheme, and

increasing a length of a solution path determined by a stored repertoire of pairs of values.

3. A method according to claim 1, and further comprising the step of varying stored pairs of the values to increase a repertoire of the robot.

4. A method according to claim 3, and further comprising the step of varying the control values which are fed into the strategy generating network by unstable dynamics of the network system and a classification of a dynamic analyzer when a new pair of values is present.

5. A method according to claim 3, and further comprising the steps of systematically generating new variants in an on-line exploratory phase by replacing a subset of the control values generated with the strategy generating network with random values, and subsequently storing said new variants.

6. A method according to claim 3, and further comprising the steps of disconnecting the networks from the sensory device and acting units in an off-line phase, replacing the captured values by random values with tunable amplitude, and selecting resulting sequences of pairs of values according to different criteria and trained to the networks of the strategy generating network and the network system.

7. A method according to claim 1, and further comprising the step of inputting output values of the plurality of further networks of the network system to the strategy generating network.

8. A method according to claim 1, and further comprising the step of feeding output values of the plurality of further networks of the network system into a dynamic analyzer which determines trained pairs of values from unknown pairs of values based on stability of the output values.

9. A method according to claim 8, and further comprising the step of giving a result of the dynamic analyzer as additional input to the strategy generating network.

10. A method according to claim 8, and further comprising the step of giving the output values of the plurality of further networks of the network system as inputs simultaneously to the strategy generating network and the dynamic analyzer.

11. A method according to claim 1, and further comprising the step of preprocessing the values fed into the network system by Kohonen-maps.

12. A method according to claim 1, and further comprising the step of preprocessing the values which are fed into the strategy generating network by at least one Kohonen-map.

13. A method according to claim 1, and further comprising the step of providing, during the training phase, selected pairs of values as learning examples to the strategy generating network and to the plurality of further networks of the network system.

14. A method according to claim 1, and further comprising the steps of delivering desired performance by taking into account inherent symmetry of an arrangement of the network system while training on the network system, guaranteeing symmetry of reciprocal interconnections between the plurality of further networks of the network system for trained pairs of values which stays stable despite fluctuations which reveal asymmetry introduced by said unknown pairs of values by inducing instability, and establishing symmetry for newly stored pairs of values while affecting previously stored patterns as little as possible.

15. A method according to claim 14, and further comprising the step of replacing the values captured by the sensory device and the control values generated by the strategy generating network during the training phase.

16. A method according to claim 1, and further comprising the step of using the capturing, generating, feeding and distinguishing steps as a module in a bigger network of such modules.

17. A method according to claim 1, and further comprising the steps of generating a fractal network architecture by recursive replacement of a single module with a pack of four interconnected modules, and interconnecting the modules so that interconnection within the pack is organized hierarchically while the connections of the pack to outside structures are organized in parallel.

18. A method according to claim 17, and further comprising the steps of keeping a repertoire of the single module and maintaining functional continuity with one of said four interconnected modules until the other modules of said four interconnected modules acquire a sufficient repertoire from said one of said four interconnected modules, and controlling composition of an output of the pack by controlling superposition weights with said one of said four interconnected modules.

19. A method according to claim 17, and further comprising the steps of communicating two of said interconnected modules with nonoverlapping subsets of the sensory device, and consequently performing different input/output relations with acting units so that division and corresponding reassembling of domains of operation define a topographic principle and are performed with regard to a periphery.

20. A method according to claim 19, and further comprising the step of receiving a feedback projection of an output of one of said two of said interconnected modules with at least one other of said four interconnected modules so that a feedback projection is not filtered by a Kohonen-map.

21. A method according to claim 17, and further comprising the steps of receiving, with one of said interconnected modules, an output of two others of said interconnected modules, projecting the output back to a united domain of the acting units and reassembling outputs of all modules in a weighted superposition.

22. A method according to claim 21, and further comprising the step of receiving, with said one of said interconnected modules, a feedback input from an output of another module in said united domain.

23. A method according to claim 17, and further comprising the steps of receiving, with one of said interconnected modules, outputs from the other modules of said interconnected modules and from a whole domain of a periphery, sending, from said one of said interconnected modules, control sets to all acting units of a topographic domain of the pack, and controlling, with said one of said interconnected modules, superposition of outputs of all modules of the pack.

24. A method according to claim 17, and further comprising the step of precluding superposition of parallel outputs of the pack of four interconnected modules if a receiving structure provides enough lines of communication to receive the output of all of said four interconnected modules in parallel.

25. A device for controlling an autonomously exploring robot comprising:

an artificial neuronal strategy generating network which generates control values for acting units of the robot according to input and trained synapse strengths,

a sensory device for capturing values from an environment which are fed into the strategy generating network as input values,

acting units which receive control values from the strategy generating network,

a short term memory for storing said input values of the sensory device and corresponding control values, which constitute a pair of values,

an artificial neuronal network system, including a plurality of further networks, which is supplied with the values captured by the sensory device and the control values generated by the strategy generating network, generates stable output values upon receiving a pair of values already learned during a training process, and generates instable output values upon receiving an unknown pair of values, output values of the artificial neuronal network system being supplied to a signal generating network, and

a dynamic analyzer for distinguishing between pairs of values learned during a training phase and unknown pairs of values on the criterion of dynamic stability of the network system, an output of which may be connected to the strategy generating network.

26. A device according to claim 25, wherein the network system includes two networks.

27. A device according to claim 26, wherein each network is connected to a preprocessing Kohonen-map, output values of each network being back to the network'"'"'s own preprocessing Kohonen-map and to the preprocessing Kohonen-map of the other network.

28. A device according to claim 27, wherein the dynamic analyzer receives the output values of the networks of the network system, and classifies the input to the network system, based on stability of the output values of the two networks, as known or unknown, classification of the stability of the output of the two networks falling into classes of local and global, go, no-go, and chaos states, the classification and output values of the networks being given as input to the strategy generating network.

29. A device according to claim 28, wherein the dynamic analyzer is a sliding register.

30. A device according to claim 28, wherein the dynamic analyzer is a Fourier-analyzing device with subsequent comparison of spectra.

31. A device according to claim 25, wherein the short term memory is designed as a sliding register.

32. A device according to claim 25, wherein the values captured by the sensory device and the control values generated by the strategy generating network converge on a preprocessing Kohonen-map which supplies its output to the strategy generating network.

33. A device according to claim 25, and further comprising equally designed interconnected modules which, together with the device, form a superarchitecture.