Hello Everyone, I am new to this forum.
I have been making AI for games for about 15 years now (as a hobby, not profession) and have moved into applying my AI projects onto robotic platforms. I am now moving to the Aibo platform due simply to how impressive it is.
The Ai concept I have developed for controlling the Aibo and having learn and react is as follows. Im looking for any and all feedback on anything here to refine the concept before moving to practical testing.
OVERVIEW :
3 Level Ai system. The High Level Ai uses emotions, needs, sensory input and direct user commands to choose High Level Functions (Activities) and generates a Queue of Middle Level Functions in an attempt to complete that activity. Middle Level Ai then executes each Middle Level Function (Ability) called by the High Level Ai generating a queue of Low Level Functions for each abiilty used. Low Level Ai executes each Low Level Function called by the Middle Level Ai and generates a queue of Direct Control Functions (Actions) to achieve that ability.
Learning or Adaptation is done via fitness based evaluation of each Function after it is run by an Ai Level. Whenever the Middle Level Ai succeeds at completing the High Level Function it stores the queue of Middle Level Functions that were used according to their fitness rating. Future Executions of the High Level function will use stored successful queues unless the "desire" to learn is high (Stimulation need)
Queues are generated randomly using the Decision Matrix to decide which functions can be chosen, and which functions should be chosen.
Whenever a function succeeds or fails, or the user inputs success or failure, the probability weights for the function tree in use (Activities, Abilites and Actions used) are modified accordingly. This Probability is used by the queue generator to choose functions to place in the queue.
DETAILS :
The Decision Matrix
Initial Functions and Preferred Links (Each Middle Level function is actually connected to every low level action. What you see here is their preferred actions)
Each Entry in the decision matrix begins with a 100% probability of being chosen. And queue's generated are completely random (and will likely fail)
Links between levels represent the probability that a lower level function will be chosen by a higher level function. Each link visible in the image has 100% probability to be chosen and all lower level functions not linked have a 10% chance to be chosen by that higher level function.
As Functions succeed or fail, and as queues succeed or fail., the probabilities of functions being chosen, and the link probability % between levels is modified appropriately.
PERSONALITY CONTROL
>Emotions (Restricted by Needs, Decay over time)
Happiness
Sadness
Fear
Shock
Excitement
Affection
Anger
>Needs (Inc/Dec over time)
Enjoyment
Rest
Stimulation
Excersize
Interaction
Isolation
Company
AI SYSTEM FEATURES
>High Level Ai
Regulates Emotions and Needs
Monitors Sensory Input and Converts to situational variables
Manages Decision Matrix
Manages Memory
Manages High Level Function Queue
Manages Queue Fitness and offload to Memory Manager
Evaluates Middle Level Ai Function fitness
>Middle Level Ai
Manages Middle Level Function Queue
Manages Queue Fitness and offload to Memory Manager
Evaluates Low Level Ai Function fitness
>Low Level Ai
Movement System
Voice System
Sound System
Stance System
Animation System
Search System
Kick System
Contextual Skit System
>Animation Player (dances etc)
Skit System (Skits are AIBO animation/sound files)
Contents of a function
Memory of Successful queues from previous runs
Need/Emotion Requirement to start
Variable/Time based Success restrictions
Need/Emotion Modification on Success/Failure
Need/Emotion Modification on Praise/Scold
List of Preferred Actions
List of Usable Animations
Preferred Higher Level Queue Position
Lower Level Queue is generated according to the above information and Decision Matrix
Fitness
Fitness of a function or queue is determined according the success/failure of actions used in the function/queue plus any praise/scold given and finally adjusted according to the amount of time taken.
Learning
Ai starts in learning mode, and will generate new queues for evaluation whenever acting autonomously until at least 50% of function memory is used. Once this is reached, Improvement takes over learning
At this stage complete queues are generated at random using the decision matrix.
Improvement
The "Stimulation" need controls when new queues are generated. New Queues are evaluated and added to memory
At this stage stored queues are mutated and evaluated if stored queues are available. If memory is full, replace lowest fitness queue with new queue if fitness is higher. Repeated failures will switch the Ai back to learning mode to generate an entirely new queue instead of mutating.
I hope that is enough information for you to get an idea of how it works
