1.10 Introspecting AI
Requirements
Modify the Teachable Bot program to permit Trainers and Admins to set AI players to conduct experiments on their own. An experiment involves two AI roles: an “explorer” which invents a hypothesis (a proposed move), and a “debater” which invents a test of that hypothesis by naming a sequence of responses opponents might make. If one player plays both roles, the experiment is called “introspection” or “self-play”, and the hypothesis is not considered “independently tested” (so biases pose greater risk). If we find that humans out-perform bots on some games (e.g. TheoryOfMind) only until bots are given the ability to conduct experiments on their own, then we will have shown that conducting experiments on one’s own is necessary to reach human-level intelligence.
Create a global constant called “Research Budget” representing the maximum number of experiments each AI can conduct per actual turn. Allow Trainers and Admins to set non-zero parameters on each AI for Introvert, Empath, and Curious, and add Explored (replacing Played), Debated, Research, Empath (EMP), Teach (TCH), and Research Speed (RS) to the player statistics for AI.
When an AI plays (rather than augments), it should decide whether to test its proposed move by generating a random number (from 0-1). If the AIs score for its proposed move exceeds a threshold of max (0, Introvert - the random number), then the AI should accept its hypothesis without further testing. Otherwise, it should debate (without user review), and keep conducting experiments until either the time or budget for its turn runs out or until it scores a move as exceeding the threshold (e.g. because the debate sufficiently boosts confidence in the original hypothesis or in an alternate hypothesis).
Learning to Debate
Record debater moves as type=”debate” (instead of type=”explore”), whether the debater predicted the move would “teach” the explorer, whether the debater predicted the move would actually be made, and whether the move did teach and was actually made. If any explore in the debate predicted a better outcome than the actual debate outcome (e.g. if the explorer ever predicted “win”, “strategic_lose”, “draw”, or “unstrategic_win” but the assigned outcome of the debate was ”unstrategic_draw”), then count all all prior debate moves of that debate as “teach”. For “debate” moves that the real opponent actually made, count them as “actual”.
Acceptance Test Plan
Test this feature on a variety of rule sets at a variety of Research Budgets and “Personality” settings.
Set a reasonable max Research Budget (maybe 3?). Fork your best non-introverted bots to make versions with the best Introvert, Empath, Curious settings. Benchmark these new bots against Random and their parents. As in 1.9 Teachable AI, define a new non-introverted bot that uses the the tournament against the introverted bots as curriculum, then benchmark it against the bots in the tournament. Can we use this strategy to avoid research for most bots (i.e. limit the use of research to the creation of curriculum)? Does it depend upon the kind of game?
Demonstrate the security advantages of Skepticism
Use an AI that never had continuous learning on but for whom the curriculum is dominant/submissive play on 4on7sq from 1.9 Teachable AI. Set Offense, Tactical, Introvert, Empath and Curious to 0.5, but Faith to 1.0. Does it exhibit submission? How long does the behavior remain stable? Try the same thing with Faith set to 0.1.
Demonstrate the value of Empath
Develop the best 4P-Blind-TTT AI you can, then generate two forks: one tuned to 4P-Blind-TTT and the other identical except with Empath set to 0. Benchmark the two new AI against each other, Random, and their parent. Compare win rates, ratings, learning curves, favoritism, and profiles.
Demonstrate security vulnerability of Empath
Use an AI that never had continuous learning on and was never given curriculum (i.e. an AI that learns entirely by conducting experiments). Set Offense and Tactical = 1.0, Faith = 0.1, Introvert and Empath = 1.0, and Curious = 0.0. Apply the same 3on15line strategy as in 1.9 Teachable AI.
Formulae
Introvert (vs Extrovert)
- \(\text{Introvert}\) :
\(1.0\) means practice on simulations as much as possible before acting; \(0.0\) means learn only via action (so as not to lose touch with “reality”)
Empath (vs Projective)
- \(\text{Empath}\) :
\(1.0\) means try to predict others’ moves based on their stats and recent behavior; \(0.0\) means expect others to do whatever you would do in their situation
Curious (vs Practical)
- \(\text{Curious}\) :
\(1.0\) means practice unexpected scenarios as much as possible; \(0.0\) means practice only expected scenarios
- \(\text{dScore}_x\) :
How much the classifier recommends moving the debate with game state \(x\)
Research Metrics
- \(\text{Explored}_{a, g}\) :
The number of matches (or experiments) of game \(g\) explored by classifier \(a\)
- \(\text{Debated}_{a, g}\) :
The number of experiments of game \(g\) debated by classifier \(a\)
- \(\text{Research}_{a, g, n}\) :
The average number of number of experiments conducted by classifier \(a\) on game \(g\) in the 100 moves ending with \(n\)
- \(\text{aCount}_{a, g, n}\) :
The number of correct “actual” predictions by classifier \(a\) among the 100 predictions of game \(g\) ending with move \(n\)
- \(\text{EMP}_{a, g, n}\) :
The F1 of classifier \(a\) for predicting the actual moves of other players on game \(g\) for the 100 moves ending with \(n\)
- \(\text{TCH}_{a, g, n}\) :
The F1 of classifier \(a\) for predicting debate moves in game \(g\) that will teach the explorer for the 100 moves ending with \(n\)
- \(\text{RS}_{a, g, n}\) :
The speed with which classifier \(a\) has conducted research on math:g in the 100 moves ending with \(n\)